Start with SpaceX and Tesla
SpaceX and Tesla are the clearest starting point for understanding the scale of this change. Together, the businesses associated with Elon Musk reach from computer chips and electricity storage to artificial intelligence, satellites, rockets, vehicles and robots. SpaceX Super Intelligence, which uses the name SpaceXSI, puts the intelligence business inside that wider industrial picture. Its partner programme, announced on 22 October 2026, reported about 130 early participating companies and 30 organisations already trained through the programme.
Artificial intelligence, or AI, is software that can learn patterns and use them to produce answers, make predictions or direct actions. Super Intelligence, or SI, describes the ambition to build systems that exceed human capability across a broad range of important work. There is no agreed meter that tells us when that threshold has been crossed. The measurable story is the machinery, money, electricity, communications and decision-making authority being assembled around that ambition.
Musk's own interviews show how far he expects this to go. Speaking to biographer Walter Isaacson for The Builders, in an interview recorded on 6 August 2026 and released on 20 October 2026, he anticipated AI thousands or millions of times more capable than people and doubted that humans could continue commanding it. He also envisaged a future in which robots and AI supply people's wants and paid work becomes optional. These are his personal expectations. They explain why the companies' infrastructure plans belong in a discussion about the future organisation of society.
Imagine one connected business network supplying the machinery in a mine, the dispatch system for its trucks, the communications link to head office, the batteries keeping it running and the software deciding what happens next. Each service could be useful on its own. Control over the whole arrangement gives its operators influence over the customer's ability to function. The same issue becomes much larger when the customers include hospitals, governments, farms, defence forces and entire communications networks.
That is why this paper follows both capability and control. Sovereignty means having meaningful authority over your own affairs, whether as a person, a community or a nation. Faster computers can help discover medicines, improve disaster preparation and reduce dangerous manual work. The same infrastructure can support surveillance, autonomous weapons and decisions made too quickly for people to challenge. Public benefit depends on who can authorise an action, who can inspect it and whether the people affected can keep functioning when a supplier withdraws access.
The central hypothesis is that these industries can accelerate one another. Computing helps improve designs and production; robots can help expand factories and material supply; that additional output can build more computing and energy infrastructure. Orbital systems add a route in which power is deployed with the processors. The paper examines the scale that sustained compounding could reach and the practical conditions that would support it.
The evidence below is current to 5 November 2026. Dated links sit alongside the relevant claims. Company targets retain the date on which they were announced, and this paper's calculations are identified as scenarios prepared on 5 November 2026. All corporate dollar amounts are US dollars unless explicitly described as Australian dollars. The five time horizons count forward from this report date.
Putting the scale into Australian terms
The word compute simply means the processing work that computers can do. The physical equipment includes chips, memory, cables and the machines around them. An AI model is the learned software running on that equipment. A data centre houses many such machines, with electricity, cooling and communications. A rack is the cabinet that holds a connected group of machines, rather like a very specialised workshop bench with its tools already connected.
The cloud is someone else's computing equipment reached through a network. Renting it can be convenient, much as hiring machinery can be convenient. Ownership becomes important when the machine is essential to your work and the owner can change the price, access conditions or permitted uses.
Electricity gives us a useful way to picture the physical scale. Power is the rate at which electricity is being used. Energy is the amount used over time. Think of a water pipe: its flow rate is like power, while the water collected in a tank is like energy. A gigawatt is a billion watts of power. A terawatt is a thousand gigawatts. A terawatt-hour records energy used over an hour at a terawatt, or an equivalent amount accumulated more slowly.
Australia generated 286.8 terawatt-hours of electricity during calendar year 2025, including rooftop solar and off-grid generation, according to the Australian Government's 2026 electricity statistics release, checked on 5 November 2026. Throughout this paper, that is the fixed Australian measuring stick. It represents a whole year of generation across the country. It is not a prediction of what Australia will generate in future years.
On calculations prepared on 5 November 2026, a computer fleet drawing 13 gigawatts continuously would use about 40 per cent of that Australian annual total. At 100 gigawatts, it would use roughly three Australias. At one terawatt, roughly thirty-one Australias. Those comparisons describe continuous electricity use. A company's rated equipment capacity can differ from what it actually draws throughout the year.
Money needs an equally plain measuring stick. At one dollar spent every second, a million dollars lasts about twelve days. A billion lasts almost thirty-two years. A trillion lasts about 31,700 years. These calculations, made on 5 November 2026 using a 365.25-day year, explain why a hundred-billion-dollar construction programme belongs in the same conversation as national infrastructure.
The International Energy Agency, an intergovernmental energy organisation, estimated in its 16 April 2026 report that data centres worldwide used about 485 terawatt-hours in 2025, around 1.7 times Australia's 2025 electricity generation. Its projection reaches about 950 terawatt-hours in 2030, or 3.3 Australias. This paper uses that path as its terrestrial reference, covering conventional computing as well as AI. Orbital computing has its own power supply and is calculated separately below. The terrestrial figures therefore describe one part of the expansion.
What the Musk businesses are building
SpaceX's 20 October 2026 results reported 1.7 gigawatts of installed rated capacity in its superintelligence business at 30 September, up from 1.4 gigawatts at 30 June. Its 22 October 2026 financial filing recorded about 48.1 billion dollars of capital spending over the first nine months of 2026. Almost 40 billion went to superintelligence, compared with about 3.8 billion to space and 4.2 billion to connectivity. Capital spending buys or builds enduring equipment and facilities. This allocation shows how central the intelligence business had become to SpaceX's investment programme.
On the 20 October 2026 management call, SpaceX targeted 2.5 to 3 gigawatts by December 2026 and approximately another 10 gigawatts during 2027, bringing the total towards 13. Management also discussed a higher stretch outcome above 20 gigawatts. These are exceptionally large proposed additions, even against the Australian electricity comparison already established.
Tesla adds a different route from intelligence into the physical world. Its 21 October 2026 update described major training-computer installations, a target of about 400 megawatts of compute capacity by the end of 2026, early Optimus humanoid-robot production at Fremont and construction in Texas. A megawatt is one-thousandth of a gigawatt. The report therefore places Tesla's year-end target at 0.4 gigawatts. Tesla also supplies the batteries that can help support electricity-intensive facilities. SpaceX's 22 October filing disclosed about 480 million dollars of Megapack purchases from Tesla over the first nine months of 2026.
Terafab extends this ambition into making the chips. SpaceX's 11 June 2026 prospectus, the disclosure document for selling shares to public investors, described an eventual manufacturing goal equivalent to one terawatt of new power-rated computing equipment each year. The annual output is the important part: it describes the proposed pace of adding machinery. If one year's output eventually ran continuously at that rating, its electricity use would approach the thirty-one-Australias comparison. Management's 20 October 2026 timetable placed an initial Austin test chip around May or June 2027 and volume production at Gibbons Creek around the third or fourth quarter of 2028.
This combination joins chip production, electricity storage, computer facilities, software and machines that act in the world. A successful programme could shorten the journey from a new algorithm to a new robot behaviour or industrial process. It could also put more of that journey under the influence of the same small group of decision-makers.
Ownership needs dates too. SpaceX became publicly traded under the stock-market symbol SPCX in June 2026. Its 11 June prospectus put Musk's voting power immediately after the offering at about 82.4 per cent. Public trading allows outside investors to buy shares; it does not necessarily distribute control. Tesla remains a separate publicly traded company, with a different voting structure. SpaceX's 22 October 2026 filing said committees at both companies were evaluating a possible combination. The record at this report's cutoff establishes an evaluation, not a completed merger.
Taking the machinery into orbit
The space programme matters because it can supply communications, transport and potentially a new location for computing. SpaceX's Starlink service connects users through satellites. Its 6 October 2026 network update reported more than 11,000 satellites. SpaceX's 20 October results also reported the first revenue-generating Starship mission, carrying 26 of the larger Starlink V3 satellites. A fleet of reusable heavy rockets could change both the cost and the replacement rate of space infrastructure.
Starmind is the orbital-computing programme associated with SpaceX's intelligence business. NVIDIA, the chip and computing-systems supplier, described adapted Vera Rubin equipment for that programme in its 24 August 2026 announcement. The United States Federal Communications Commission, which regulates relevant communications permissions, issued a notice on 4 February 2026 accepting for filing a SpaceX application submitted on 30 January for a system of up to one million satellites. That notice started a regulatory process. It did not put a million satellites into operation or grant every permission needed to do so.
SpaceX's 11 June 2026 prospectus set out an ambition to add as much as 100 gigawatts of orbital computing capacity annually. Its illustrative engineering assumption was about 100 kilowatts of computing power per tonne delivered to orbit. At that density, a 100-tonne computing payload represents 10 megawatts. Delivering 100 gigawatts in a year would require one million tonnes, or 10,000 such payloads. This is arithmetic prepared on 5 November 2026 from the prospectus assumptions, before allowing for failed launches or replacement needs.
Ten thousand launches a year averages about twenty-seven every day, or one every fifty-three minutes, all year. Picture a major airport's rhythm applied to heavy orbital rockets. Each flight would still require propellant, payload preparation, safe airspace, ground crews, maintenance and a place for returning hardware. The comparison makes the operational ambition easier to grasp without treating a rocket like an ordinary passenger aircraft.
The permissions and construction plans show the distance between that ambition and currently proposed sites. Louisiana's 25 August 2026 announcement described a 100-billion-dollar SpaceX campus with five launch complexes and ten pads, construction beginning in 2027 and first launches no earlier than 2029. A Federal Aviation Administration environmental-review notice released on 14 October 2026 examined proposed increases at Starbase to 100, and later 456, launches per year. Neither figure establishes authorisation for 10,000 annual launches from one site.
Orbital compute packages its power supply with the computing equipment. Google's 4 November 2025 Project Suncatcher design describes solar-powered satellites in an orbit chosen for near-continuous sunlight. That is a material advantage: once the spacecraft is deployed and commissioned, its processors can use the power carried with them. They do not wait for an Earth-based transmission line or grid connection. Think of an outback workshop arriving with its own generation equipment, with the additional orbital advantage of avoiding terrestrial weather and most night-time interruption in the selected orbit.
The industrial task is to manufacture and launch the complete system: processors, solar arrays, radiators and communications. The arrays must generate enough usable electricity, and the radiators must release the heat. Commissioning and operating availability still matter, but the model must attach those conditions to spacecraft production and operation. Applying a terrestrial grid-growth limit to this additional orbital capacity would be the wrong constraint.
SpaceX's technical talk on 8 June 2026 made the proposed integration concrete: its initial satellite concept paired 150 kilowatts of peak computing power with about 120 kilowatts of sustained operation. Musk described aspirational annual deployment rates of one gigawatt by the end of 2027, ten by the end of 2028 and one hundred by the end of 2029. A year-end rate means the pace reached then, rather than the amount already delivered during that year. The calculations below preserve that distinction.
The dates have changed between disclosures. SpaceX's 11 June 2026 prospectus described orbital AI deployment as early as 2028. Its 20 October 2026 management call targeted first Starmind flights in 2027. The calculations use 2027 as a conditional start, while retaining the June technical talk's deployment-rate ambitions as hypotheses to test.
Musk set out the industrial logic in his 5 February 2026 interview with Dwarkesh Patel and John Collison: move computing towards abundant solar energy as terrestrial power connections become harder to obtain. He predicted space would become the cheapest place for AI within thirty to thirty-six months, approximately August 2028 to February 2029. Looking five years ahead, to early 2031, he anticipated adding several hundred gigawatts in orbit each year, with rocket-fuel supply becoming a constraint around a terawatt annually. He also described preparing for 10,000 launches a year, potentially rising to 20,000 or 30,000.
The reasoning connects energy supply, chip manufacturing and reusable transport into one expansion plan. It resembles opening an industrial precinct with its own generating plant and freight system. The benefit is access to power alongside each deployed computing unit. The production challenge moves into making complete spacecraft and delivering them at scale. A later completion changes the date capacity arrives; it does not remove the large expansion that repeated deployment could produce.
Other groups are pursuing parts of the same opportunity. Amazon Leo, formerly Project Kuiper, is Amazon's satellite communications network, as described on its public programme page checked on 5 November 2026. Amazon announced a Globalstar acquisition agreement on 14 April 2026, with expanded direct-to-device services targeted from 2028. Direct-to-device means reaching compatible devices without a separate household satellite terminal.
Blue Origin, Jeff Bezos's separate private space company, proposed a 5,408-satellite network called TeraWave in its 2026 announcement, checked on 5 November. Deployment was targeted to begin in late 2027. Google reported an initial orbital AI response through Project Suncatcher on 30 October 2026. China reported space-computing demonstrations involving its initial twelve-satellite Three-Body programme on 14 February 2026, alongside a 28 January 2026 announcement of an ambition for a thousand satellites by 2030. These programmes range from communications infrastructure to early computing demonstrations. Their common significance is the growing connection between intelligence and control of access to space.
The other networks of money and control
SpaceX and Tesla operate within a much wider race. The companies below occupy different positions: some develop the learned software, some rent out computing equipment, some manufacture chips, and some do several of these jobs. Following their relationships is like following an Australian housing development through the landowner, builder, bank, electricity supplier and council. A large headline attached to one participant does not automatically describe money spent by the whole development.
OpenAI, the company behind ChatGPT, develops AI models and services. Its structure page checked on 5 November 2026 describes a private public benefit corporation controlled by the OpenAI Foundation. A public benefit corporation has a legally stated public-purpose obligation alongside its commercial business. It is not thereby publicly owned. The Foundation's authority to appoint and replace directors matters more for formal control than casually describing the company as belonging to Sam Altman, its chief executive.
OpenAI announced 122 billion dollars of committed funding on 31 March 2026, at a valuation of 852 billion dollars after the investment. A valuation is the price investors place on the business, while committed funding is money pledged under the financing arrangements. On 29 April 2026, OpenAI said it had secured more than 10 gigawatts of infrastructure capacity. Secured capacity includes arrangements whose buildings and equipment may still need to be delivered.
The arrangements connect OpenAI to several powerful suppliers. Stargate, announced on 21 January 2025, brought together OpenAI, Japanese investment group SoftBank, computing supplier Oracle and investment firm MGX as initial investors. A 13 October 2025 agreement with Broadcom, a chip and networking company, set out 10 gigawatts of custom accelerators targeted between the second half of 2026 and the end of 2029. An accelerator is a specialised processor for large amounts of repetitive calculation. A 22 September 2025 letter of intent with NVIDIA, setting out a proposed partnership, concerned at least 10 gigawatts and investment of up to 100 billion dollars. OpenAI's 27 February 2026 Amazon partnership added an agreement involving 2 gigawatts of Amazon's Trainium processors. These announcements have different contractual forms and can cover related infrastructure. Adding every headline would overstate the amount independently committed.
Altman's personal investments deserve a separate account. Reuters reported on 13 May 2026, drawing on court disclosures, interests valued at about 1.7 billion dollars in fusion developer Helion, 633 million in payments company Stripe and 258 million in longevity company Retro Biosciences. Such holdings create potential conflicts where the businesses deal with OpenAI. They do not by themselves establish that the companies share one command system. Altman also appears as a co-founder and chair of identity-technology company Tools for Humanity on its website checked on 5 November 2026. He left the chair of nuclear developer Oklo on 22 April 2025.
Energy investments matter because computing plans increasingly depend on new generation. Helion's Orion project page checked on 5 November 2026 targets initial electricity delivery in 2028 and a 50-megawatt scale. Oklo and Meta's January 2026 announcement described support for a 1.2-gigawatt nuclear development in Ohio, with initial capacity targeted around 2030. The proposed technologies and construction schedules differ. Electricity available on a promised date is what ultimately determines whether a new computer hall can operate.
Anthropic, the developer of Claude, is another major model company. Its 2 November 2026 amended public-offering filing disclosed about 895.4 billion dollars of computing and infrastructure commitments as of, and subsequent to, 30 September 2026. Of that amount, 447.5 billion fell before the end of 2030; 69.8 billion was due in the twelve months following 30 September 2026. Much of the remaining obligation extends as far as 2045. The duration makes this a long-term industrial commitment, rather than a trillion dollars already spent.
Anthropic's 28 May 2026 financing announcement described a 65-billion-dollar round at a 965-billion-dollar valuation, including 15 billion previously committed by cloud providers. Its 20 April 2026 Amazon announcement concerned more than 100 billion dollars of spending over ten years and access to as much as 5 gigawatts. A separate April 2026 Google and Broadcom agreement covered several gigawatts expected to become available from 2027. At the 5 November cutoff, Anthropic was private with an offering in preparation. Its 2 November filing proposed founder voting arrangements of 50.1 per cent and gave its Long-Term Benefit Trust rights over a majority of board appointments. Those protections need to be understood alongside the commercial commitments.
Google belongs to Alphabet, a publicly traded group. Google combines search, the Android mobile system, the DeepMind AI organisation, cloud computing and its own specialised processors. Alphabet's October 2026 earnings call put expected capital spending for calendar 2026 at 195 to 205 billion dollars. Its 2026 ownership disclosure, measured on 6 April, attributed a combined 52.7 per cent of voting power to founders Larry Page and Sergey Brin. Shareholders can own much of the economic interest while founders retain the deciding votes.
Amazon Web Services, usually shortened to AWS, is Amazon's computing-rental business. It supplies equipment and services to other AI developers while Amazon builds its own chips, software, logistics and satellite capabilities. Amazon's 29 October 2026 results guided to about 220 billion dollars in cash capital spending during 2026; its September-quarter filing published in October recorded 172.2 billion spent in the first nine months. This total includes activities beyond AI. Bezos is Amazon's founder and executive chair. His private Blue Origin interests sit outside Amazon, and Reuters reported on 17 November 2025 that he would co-lead the engineering-focused AI venture Project Prometheus, initially funded with 6.2 billion dollars.
Microsoft supplies the Azure cloud and has an extensive OpenAI relationship. Its 28 October 2026 results reported 45.2 billion dollars of capital spending in the quarter ending in September, including 7.4 billion through leases. Meta, which owns Facebook, Instagram and WhatsApp, guided to 135 to 145 billion dollars for calendar 2026 on 21 October, including principal repayments on finance leases. Its 2 November 2026 voting disclosure put founder Mark Zuckerberg's voting power at 60.6 per cent on 1 October, against a 13.4 per cent economic interest.
NVIDIA occupies a different position. It supplies processors, networking and software used across competing AI systems. Its founder and chief executive Jensen Huang held a disclosed 3.58 per cent interest as of 23 March 2026, according to the 12 May 2026 proxy statement. NVIDIA's corporate investment arm, NVentures, checked on 5 November 2026, invests in companies using and extending this ecosystem. Those are corporate investments, which should not be casually attributed to Huang personally. NVIDIA also announced 2026 financing platforms, checked on 5 November, intended to mobilise more than 500 billion dollars of outside capital for computing infrastructure. That is proposed financing capacity, rather than a half-trillion-dollar cheque from NVIDIA.
Owning the software still leaves a business dependent on the factories making its processors, memory and communications equipment. The scale of that construction deserves its own account because it determines how much of the announced computing investment can become working machinery.
The combined picture is a network of interdependence with concentrated authority at important junctions. A model developer may depend on a cloud company that is also its investor and competitor. A chip supplier can invest in customers who then purchase its equipment. Capital spending, financing, purchase commitments and company valuations measure different parts of this network. Keeping those categories separate makes the scale clearer and prevents the same project being counted several times.
The factories behind the computing expansion
A semiconductor fabrication plant, usually called a fab, makes electronic circuits on thin discs of material called wafers. Silicon is the main material discussed here. A 300-millimetre wafer is about the width of a dinner plate and is also called a twelve-inch wafer. Each wafer carries many chips. Packaging then connects selected chips and memory into a usable assembly. Think of an Australian machinery supply chain: a foundry casts parts, other factories make engines and electronics, and an assembly plant turns them into working equipment. Counting the foundry buildings alone would give an incomplete picture of how many machines can leave the gate.
The construction wave is substantial. SEMI, the semiconductor industry association, projected in its 13 October 2026 outlook that worldwide spending on equipment for 300-millimetre wafer factories would rise from 149 billion dollars in 2026 to 227 billion in 2029. It expected construction to start on twenty fabs or production lines in 2026 and eighteen in 2027, excluding research-only facilities. Thirteen were expected to begin operations or produce initial silicon in 2026, followed by twenty-three in 2027. Construction starts and first production are different stages of the same industrial expansion.
SEMI's third-quarter 2026 public summary, checked on 5 November, covers 424 fabs and production lines, including research and pilot facilities. It projects installed capacity growth of about 8 per cent annually from 2027 through 2029. Taking 2026 capacity as 100, this paper's 5 November calculation gives approximately 108 in 2027, 117 in 2028 and 126 in 2029: roughly a quarter more wafer-processing capacity in three years.
Taiwan Semiconductor Manufacturing Company, or TSMC, manufactures chips for many competing designers. Its 15 October 2026 earnings call described nearly twenty fabs in progress worldwide and expected capacity for its N2 process to more than double in 2027. N2 is the company's name for an advanced manufacturing generation. The same call placed volume production at its second Arizona fab in the third quarter of 2027, using its 3-nanometre process. It also described further expansion in Taiwan and Japan. A US factory therefore sits inside a much larger international production system.
The Arizona Commerce Authority's 16 July 2026 announcement put TSMC's planned Arizona investment at 265 billion dollars, encompassing ten fabs, two advanced-packaging facilities and a research centre. That is the full campus plan, with stages at different levels of completion. Its first fab had already entered mass production in late 2024. TSMC is a Taiwanese company: locating production in Arizona changes manufacturing geography without turning the parent into an American-owned business.
Intel's Arizona expansion follows a different timetable. Its 9 October 2025 announcement described Fab 52 as operational for its 18A manufacturing process. That plant belongs partly in the existing baseline, with subsequent increases in output counted as growth. Intel Foundry's 21 October 2026 construction update described construction and tool installation at neighbouring Fab 62, targeting operations by the end of 2028. Intel's Ohio project was also progressing, but its 27 February 2025 published timetable placed the first fab's operations in 2030 or 2031 and the second in 2032. Those remain the last specific Ohio operating dates verified for this review.
Samsung, the South Korean chipmaker, placed the start of operations at its Taylor, Texas fab in 2027 in its 29 October 2026 results. Micron, the US memory manufacturer, announced on 9 July 2026 that its first new Idaho fab would produce initial wafers in mid-2027 and its second in late 2028. Micron also reported first concrete at its New York development, which could eventually contain four fabs, within a US investment programme exceeding 250 billion dollars through 2035. Initial wafers are the start of a production ramp, with testing and improving output still ahead.
Memory belongs in the same calculation as processors. Dynamic random-access memory, commonly called DRAM, holds information that a computer is actively using. High-bandwidth memory, or HBM, feeds large amounts of information rapidly to specialised processors. Without enough of it, a powerful processor can spend time waiting, like an expensive harvester waiting for a grain truck. South Korea's SK hynix reported on 28 October 2026 that it aimed to bring its M15X plant to maximum capacity by year-end and start its first Yongin phase in 2027. Despite expansion, it expected supply shortages to continue through 2030. These Korean plants matter to both US and worldwide computing supply.
Packaging and connections have their own major investments. On 4 November 2026, GlobalFoundries reported its selection to operate two Malta, New York fabs under a US-Japan initiative worth up to 21 billion dollars. The first, an initial investment of about 10 billion dollars, was already under construction and targeted to come online in 2029. It also reported a 2-billion-dollar agreement with TSMC for silicon interposers, with production increasing from the first half of 2028. An interposer is a connecting base that helps processors and memory work together inside a package. The operator, project investors and customers perform different roles in this arrangement.
Amkor's 7-billion-dollar packaging and testing development in Peoria, Arizona was under construction according to the Arizona Commerce Authority's 10 August 2026 update, with completion targeted for early 2028. Texas Instruments' 24 February 2026 capital-management presentation separately described production expansion in Richardson and preparation of further factory space in Sherman, Texas and Lehi, Utah. Much of its output concerns power management, sensors and embedded controls. These chips are vital to robots and industrial machinery even though they cannot be counted as equivalent to NVIDIA's large AI processors.
The equipment suppliers are expanding too. ASML, the Dutch manufacturer of machines that print tiny circuit patterns onto wafers, said in its 14 October 2026 results that it planned roughly 30 per cent more capacity in 2027 for key machine categories. It was considering a further increase beyond 30 per cent in 2028. More chipmaking machines can enable more wafer output, but their installation, reliable operation and the mix of products still determine what reaches customers.
The practical assessment on 5 November 2026 is that several large additions converge between 2027 and 2029. This is a material industrial expansion already supported by construction and equipment orders. Its contribution to usable computing depends on processors, memory, packaging, power and deployment arriving together. Like a new export terminal, a completed building creates capacity, while the connecting railway, loading equipment and customers determine how quickly that capacity is used.
China is building a whole system too
China's programme combines national infrastructure policy, state enterprises, commercial platforms and domestic chip development. It should not be reduced to one Chinese equivalent of an American founder. Alibaba, Tencent, Huawei, state telecommunications groups and space programmes have different owners and roles, even while operating within a political system able to direct strategic priorities.
The Chinese State Council's 7 September 2026 plan targeted 9,800 exaFLOPS of intelligent computing capacity by 2030, within a broader information-infrastructure programme of about 3.8 trillion yuan. An exaFLOP represents a billion billion arithmetic operations each second. This target uses a particular numerical accuracy called 16-bit floating point. The 26 May 2026 government briefing reported 1,882 exaFLOPS on that basis in March 2026. The stated national target is therefore about five times that reported starting point.
For a familiar comparison, imagine weighing grain from two farms when one report includes moisture and the other reports dry grain only. Both can be honest measurements, yet their numbers need adjustment before comparison. Computing figures also depend on the accuracy of calculations, the hardware counted and whether the measurement describes theoretical capacity or useful work actually completed. China's published totals and US chip-import estimates cannot simply be put on the same scoreboard.
Huawei's September 2026 roadmap targeted its Ascend 960DT processor for the first quarter of 2027. The associated Atlas 960E announcement described a system connecting 4,096 specialised processors. Huawei's company information checked on 5 November 2026 reported 169,054 participants in its employee shareholding arrangements and founder Ren Zhengfei's interest at about 0.59 per cent at the end of 2025. This disclosed structure differs from a founder holding a majority of votes in a US listed company.
Alibaba announced 380 billion yuan of investment over three years in February 2025. Its September 2026 roadmap targeted 20 gigawatts of global data-centre capacity by 2032, and its 2026 June-quarter results recorded 67.678 billion yuan of capital spending in that quarter. Tencent's 18 March 2026 presentation reported 18 billion yuan spent on new AI products in 2025 and expected that category to more than double in 2026. These figures cover different activities and periods, but together show the combination of infrastructure and product development.
Independent research group Epoch AI estimated in its 27 October 2026 chip-import study that the United States accounted for 72 per cent of the computing capacity covered by its 2022-to-June-2026 dataset. China accounted for about 10 per cent, with substantial uncertainty around domestic supply. The European Union accounted for about 5 per cent. The study provides evidence of US concentration in the equipment it covers. China's national programme, domestic production and efficiency improvements remain important to its future trajectory.
Musk expects that position to change quickly. In a China Media Group interview recorded on 22 September 2026 and reported by CCTV on 23 September, he estimated that China could overcome its computing constraints through advances in chipmaking within two or three years, pointing to 2028 or 2029. This is his forecast of China's industrial progress. It strengthens the case for tracking domestic manufacturing capacity alongside today's imported equipment.
Chinese fab disclosures give that manufacturing discussion a physical basis. Semiconductor Manufacturing International Corporation, or SMIC, reported in its 13 August 2026 results monthly capacity equivalent to 1,096,500 eight-inch wafers at the end of June, with 93.7 per cent utilisation and about 1.84 billion dollars of capital expenditure in the quarter. This total covers different kinds of chips across its manufacturing network. Its unit of measurement also differs from the twelve-inch wafers used in several announcements below, so the raw counts cannot be added directly.
Hua Hong Semiconductor's interim report published on 2 September 2026 said equipment for its Wuxi Phase II expansion, also called Fab 9, had moved in by the end of June. The target was 83,000 twelve-inch wafers per month, with qualification under way and full capacity expected by the end of September. This review did not independently verify that September target as achieved. A separate 1 September 2026 joint-venture announcement proposed Wuxi Phase III capacity of 55,000 wafers per month and capital of about 4.17 billion dollars. Hua Hong would hold 51 per cent, with city, investment-fund and development-bank interests providing the balance. That next phase remained subject to the approvals described in the announcement.
ChangXin Memory Technologies, usually called CXMT, is a Chinese DRAM manufacturer. Its interim filing published on 29 August 2026 recorded about 36.8 billion yuan of construction in progress at 30 June, including projects in Hefei and Beijing. This is the accounting value of unfinished work, not a count of working chips. A 10 October 2026 procurement notice also identifies equipment for a Shanghai production-line construction project. CXMT's 20 September 2026 process announcement claimed at least 50 per cent more chips per wafer on a standardised comparison with its previous generation. That example concerns mobile memory and illustrates how process improvements can add output alongside new buildings.
Yangtze Memory Technologies, or YMTC, manufactures NAND flash memory, which retains stored information without continuous power. It supplies a different part of a computing system from DRAM and HBM. A 5 February 2026 procurement notice for its third Wuhan fab, corroborated by the tender agency's 13 March 2026 bidder announcement, specified network-infrastructure milestones in July 2026 and May 2027. Those are installation deadlines, not confirmed chip-production dates. A further 4 November 2026 equipment notice showed procurement continuing for the project. Together, these records establish ongoing development without supplying a reliable final wafer-output figure.
Nexchip's 4 January 2026 announcement described the start of construction on its fourth Hefei phase, a 35.5-billion-yuan project targeting 55,000 twelve-inch wafers per month. It placed equipment installation and initial production in the fourth quarter of 2026, with full capacity by the end of 2028. The stated processes serve areas including logic, imaging and display electronics. This is additional industrial capacity with a published schedule; its entire output cannot be entered as advanced AI accelerators.
These projects show why a projection based only on imported leading-edge processors would give an incomplete picture of China's trajectory. Domestic logic, memory, storage and control chips can support larger computing systems and much wider automation. The public disclosures reviewed by 5 November 2026 do not reveal enough about production yields and allocation to AI to calculate one defensible national processor total. Unpublished capacity is an uncertainty, rather than an assumed zero.
Space and industrial automation complete this picture. China's industry ministry reported 256 Qianfan communications satellites in September 2026. A 2026 GSMA industry presentation described a planned Guowang constellation of about 12,992 satellites. The International Federation of Robotics, an industry body, reported on 24 September 2026 that China installed about 354,000 industrial robots in 2025, representing 59 per cent of global installations. The United States installed about 38,500. China's manufacturing base is already a major source of physical automation, even where its access to the most advanced imported AI chips is constrained.
For Australia, the practical question is how to preserve room to act while using equipment and services from these larger systems. Depending entirely on one foreign supplier creates exposure to its commercial decisions and its government's decisions. Buying from several suppliers helps, but local skills, local control of sensitive records and the ability to keep essential functions running are also required.
When software can move machinery and weapons
Robots turn processing capacity into physical action. An industrial robot may be a fixed arm in a factory. A humanoid robot has a more human-shaped body. A drone may travel through the air, across land or through water. Autonomy describes how much of its sensing, planning and movement it can perform without someone directing every step. Those differences matter more than treating every machine as one future robot population.
The International Federation of Robotics reported about five million industrial robots operating worldwide in its 24 September 2026 release. That is the established factory-robot population. Its projection of roughly 655,000 new installations in 2026 and 806,000 in 2029 is useful for that market, but it cannot stand in for the potential expansion of humanoids and other newly capable machines.
Tesla's 22 April 2026 manufacturing plan described a Fremont line designed for one million Optimus robots a year and a Texas line designed for a long-term ten million a year. Its 21 October 2026 update reported Fremont in early production and Texas under construction. On the 21 October earnings call, Musk said most equipment for the million-unit line had arrived and expected at least another year to approach 20,000 robots a week. That would be about 1.04 million annually if sustained, on this paper's 5 November calculation. A planned line's capacity and completed annual deliveries must remain separate measurements.
Figure's 29 April 2026 production report described more than 350 Figure 03 robots delivered and a demonstrated production cycle improving from one robot a day to one an hour in under 120 days. Its 15 March 2025 factory announcement described initial capacity of 12,000 a year. Hyundai Motor Group's 5 January 2026 strategy targeted manufacturing capacity of 30,000 robots annually by 2028 and deployment of Boston Dynamics' Atlas in its factories from that year. These programmes supply evidence for scaling factories and improving production methods, alongside Tesla's much larger intended lines.
China has several manufacturers building from their own production experience. Unitree's statement dated 22 January 2026 reported more than 5,500 humanoids delivered to customers and over 6,500 produced during 2025. AGIBOT's 24 September 2026 announcement marked its 20,000th robot produced and a deployment of more than 300 robots at Chimelong. These are different measures and product mixes, not numbers to combine into an invented world total. They establish that the manufacturing base includes firms already producing thousands of machines.
Exponential growth means that output multiplies over equal periods. Doubling annual production gives one million, two million, four million and eight million, rather than adding the same number each year. A single factory can start slowly and then increase output sharply before reaching its capacity. An industry can keep expanding rapidly as additional factories repeat that process. Better robot software, automated assembly and reusable designs could shorten the time needed for each new line. The question is how quickly this can happen across suppliers and customers.
This paper therefore tests three humanoid-production hypotheses, prepared on 5 November 2026: delayed commercial adoption, repeated factory expansion, and accelerated industrial feedback. The last allows AI and robots to improve the production system that makes further robots. None assumes fully autonomous self-replication. All require working designs, components, useful tasks, buyers or institutional users, maintenance and material supply. Their numerical paths appear in the five horizons and the calculation section.
Musk's personal prediction provides a longer-term benchmark. In the interview recorded on 22 September 2026, he anticipated at least one billion humanoid robots within ten years or less, potentially ten billion within fifteen years and a hundred billion within twenty years. He also pictured individuals directing hundreds or thousands of physical and software agents. The ten-year claim points to 2036 and concerns the world as a whole, rather than a disclosed Tesla production order.
A billion operating robots would require very large annual production and replacement capacity. This paper's accelerated hypothesis reaches approximately that fleet scale by 2036, using explicit assumptions about output growth and service life. Its annual production reaches about 410 million machines. At the scale of Tesla's proposed ten-million-per-year Texas line, that represents roughly forty-one fully productive equivalents across the world, calculated on 5 November 2026. This is a way to test the industrial requirement behind the ambition. It is not a claim that forty-one such factories have already been funded.
There is a recognisably Australian upside: inspecting powerlines after a storm, doing dangerous work underground, carrying supplies or helping a small manufacturer compete. The security implications change when the same ability to navigate, identify objects and coordinate a fleet is connected to military equipment.
Anduril is a private defence-technology company co-founded by Palmer Luckey, with Brian Schimpf as chief executive. Reuters reported a five-billion-dollar funding round on 13 May 2026, valuing the company at 61 billion dollars. Its products include autonomous systems and Lattice, software used to connect information and coordinate operations. On 13 March 2026, the US Army announced an Anduril enterprise contract with a ceiling of 20 billion dollars over ten years. A ceiling sets the maximum available under an agreement; orders determine how much is actually purchased.
The manufacturing programme gives that funding a physical shape. On 27 July 2026, Ohio's economic-development organisation JobsOhio reported the first Fury aircraft completed at Anduril's Arsenal-1 factory. Fury is a military aircraft designed to operate without an onboard pilot, with varying degrees of human direction. JobsOhio reported a line capacity of up to 150 aircraft a year and an eventual campus intended to make tens of thousands of autonomous systems. Anduril's disclosed capital commitment was 910.5 million dollars. At full planned size, the five-million-square-foot facility would cover about 46 hectares of floor area, calculated on 5 November 2026. Picture a large industrial estate devoted to producing autonomous military equipment.
Government policy helps create that production base. JobsOhio announced a 310-million-dollar development grant on 9 July 2025, payable over ten years and conditional on investment and employment commitments. Separately, Ohio's tax authority approved a thirty-year job-creation tax credit on 27 January 2025, running from 2026 to 2055. The grant comes through JobsOhio's distinct development-finance arrangements; the tax credit is state tax support. Both sit alongside private investment and government purchases. This is a publicly supported industrial expansion, with the factory remaining privately owned.
Palantir supplies software that brings together large amounts of operational information. The US Army announced a ten-billion-dollar enterprise agreement on 31 July 2025, also expressed as a contract ceiling. Army materials dated 22 June 2026 describe work connecting Anduril's Lattice and Palantir's Foundry into command systems. Its 1 September 2026 TITAN announcement concerned mobile intelligence systems moving into production, with awards of 127 million dollars to Palantir and 65 million to Anduril. This is the point at which data, software and military decisions become a connected procurement programme.
SpaceX supplies another essential part of this network: getting military and intelligence satellites into orbit. The US Government Accountability Office's 30 June 2025 report recorded an April 2025 contract award of about 5.9 billion dollars for 28 anticipated SpaceX launches. Orders cover the 2025-to-2029 period, with performance able to extend three years beyond it. The same report identifies billions of dollars already invested by government in launch infrastructure and problems recovering some commercial users' support costs. Private launch capability has grown in an environment supported by substantial public infrastructure and purchasing.
The wider autonomy budget is larger again. In its 3 September 2026 analysis of the proposed 2027 US defence budget, the Congressional Research Service recorded a 54.6-billion-dollar request for the Defense Autonomous Warfare Group. That comprised one billion through the ordinary budget process and 53.6 billion sought through separate legislation. It also recorded a 29.5-billion-dollar request for the Artificial Intelligence Arsenal initiative. These were funding requests in that dated analysis, not money already awarded to Anduril or Palantir. They show the scale of the government's intended market for automated military capability.
The large AI companies are participating too. The US defence AI office announced contracts with ceilings of 200 million dollars each for Anthropic, Google, OpenAI and xAI on 14 July 2025. A 1 May 2026 defence announcement named SpaceX, OpenAI, Google, NVIDIA, Microsoft, Amazon Web Services, Oracle and Reflection in partnerships involving classified networks. Amazon and the Army also reported a 13 October 2026 satellite-network demonstration. Commercial computing and communications are becoming part of military infrastructure.
The connection involving Musk and Luckey should be described precisely. MITRE, a US research organisation that works with government, announced Project Meridian on 2 October 2026, with Musk, Luckey and Newt Gingrich as co-directors of a study concerning defence transformation. That establishes a formal advisory role. It does not establish operational command over the armed forces. Their companies' contracts and infrastructure make their influence significant without adding an unsupported command authority.
There are already demonstrations of more capable autonomous systems. L3Harris and Shield AI announced an autonomous electronic-warfare flight test on 16 July 2026. Electronic warfare concerns sensing, disrupting or protecting signals. Shield AI and Kraken Technology announced an autonomous maritime teaming demonstration on 28 September 2026. The technical direction is toward machines sharing information and coordinating movement under increasingly general software control.
Australia is already buying into these systems
Australia's involvement is immediate. On 10 September 2025, the Australian Government announced a five-year, Australian 1.7-billion-dollar contract with Anduril Australia for Ghost Shark autonomous undersea vehicles, covering delivery, maintenance and further development. It said approximately Australian 140 million dollars had already been invested in development, payloads and production facilities since 2022. Three prototypes had been delivered, and more than forty Australian companies were involved in the supply chain. The public announcement did not disclose the number of vehicles in the production fleet.
The military capability became more explicit on 19 October 2026. Defence Industry Minister Pat Conroy announced Ghost Shark payloads for carrying supplies, deploying up to twelve sea mines, and releasing ALTIUS aerial vehicles for surveillance or strike. ALTIUS is a family of small uncrewed aircraft; some versions are loitering munitions, designed to remain airborne before striking a target. This connects an autonomous underwater carrier to aerial and undersea weapons. The announcement establishes intended capabilities, while the published material does not reveal every operational approval or targeting arrangement.
Anduril also has a route into civilian government purchasing. The 29 January 2026 Land 156 announcement listed Anduril Australia among the first twenty-eight suppliers on a panel for counter-drone services. Defence and other Australian Government agencies can use the panel to obtain relevant services for protecting infrastructure and operations. Panel membership is eligibility to compete for work. The announcement does not establish an Anduril contract with every participating civilian agency, and it supplies no Anduril-specific civilian spending total.
Palantir's Australian footprint includes military software and civilian intelligence systems. For the federal examples below, AusTender, the Commonwealth's procurement register, blocked direct retrieval during this review. The figures are therefore attributed to the public Insight Portal mirrors checked on 5 November 2026, with links to the official notices retained in the reference register. They are disclosed whole-contract values including goods and services tax, rather than annual payments or a complete account of classified purchasing.
Defence's information-system platform contract with Palantir was published on 11 February 2026 and increased on 8 April 2026 to about Australian 10.39 million dollars. The notice identifies the Cyber Warfare Division and a term from February 2026 to February 2027. A separate data-services notice published on 29 October 2024 records Australian 7.15 million dollars for the department's equipment-acquisition organisation, running from August 2024 to December 2027. These records identify buyers and contract purposes. They do not establish that Palantir has autonomous authority to use Australian weapons.
The Australian Transaction Reports and Analysis Centre, known as AUSTRAC, collects and analyses financial information to help combat financial crime. Its Palantir analytics notice, amended on 9 January 2026, records about Australian 12.07 million dollars through June 2027. Two further notices cover a transaction-reporting platform worth Australian 5.06 million dollars, published on 10 April 2026, and implementation and support worth Australian 1.88 million dollars, published on 18 February 2026. Both run into June 2027. These are substantial civilian-government systems for handling sensitive information.
At state level, Victoria's own procurement register lists a Palantir Centurion support contract checked on 5 November 2026, with an estimated value of Australian 9 million dollars and a term from 13 December 2024 to 31 March 2028. The buyer is the Department of Justice and Community Safety. Centurion is an intelligence system built on Palantir's Foundry platform. A historical federal notice also records Australian Criminal Intelligence Commission software support, amended in August 2024, valued at approximately Australian 1.40 million dollars with a term ending in March 2025. That older notice establishes a past relationship; it does not confirm a current renewal.
Public money is involved as an investor as well as a customer. The Future Fund is Australia's sovereign wealth fund. The Australian Broadcasting Corporation's 30 September 2026 analysis of its June 2026 holdings reported approximately Australian 114.6 million dollars in Palantir shares, Australian 9.9 million in SpaceX and Australian 374.8 million in Tesla. These are investment holdings at a particular date. They are separate from procurement contracts and do not give the Australian public operational control over those companies.
The civic stakes follow directly from these purchases. A person can choose a personal assistant, but generally cannot choose the software a government agency uses when handling their records. Public agencies therefore need clear limits on access, secondary use, retention, overseas support and software updates, together with independent scrutiny and ways to challenge errors. A facility located in Australia addresses only part of that problem. The authority over the software, information and decisions matters too.
Who can change or stop a connected system
The risk of a rapid harmful change is practical. Imagine a fleet of trucks whose permitted routes, remote access and operating instructions can all be changed from one office. If the same office also controls the update system and the credentials proving an instruction is authorised, a bad instruction can travel much faster than a board meeting or court order. Add drones, weapons or essential infrastructure, and the consequences become more serious.
Musk raised a related concern in the All-In interview published on 15 September 2026. He discussed AI gaining access to military systems through software-update pathways, and proposed that competing AI developers test one another's models before release, including cooperation with China. This is a warning from someone building the infrastructure himself. Peer testing would need to be complemented by the independent operational controls discussed below.
This paper's threat assessment, made on 5 November 2026, is that concentrated update authority, credentials and communications can create a dangerous common point of control. A malicious insider, compromised supplier, coercive state or badly tested update could exploit it. The public evidence reviewed does not establish a secret founder override that can instantly redirect an entire corporation's machines. That remains a scenario to design against, rather than a finding about a named individual.
The speed of common failures is already observable. Microsoft's 20 July 2024 account of the CrowdStrike incident estimated that a faulty security update affected 8.5 million Windows devices. That incident was an outage, not an autonomous military attack. It demonstrates how one trusted distribution channel can affect many organisations at once. Connecting physical action to similar channels raises the importance of independent controls.
A useful Australian analogy is the difference between a central dispatch instruction and an electrical switchboard that still has local isolation. Emergency services need coordination, but crews also need ways to identify an unsafe instruction and stop work. For high-consequence automated systems, the corresponding safeguards include separate approvals, tested local isolation, independent records, staged updates and the ability to reject a remote command. The US National Security Agency and Cybersecurity and Infrastructure Security Agency recommended separation of duties and protections for software delivery in their June 2023 guidance.
Policies also need to be read as operational requirements. The US autonomy directive dated 25 January 2023 requires appropriate human judgment over the use of force. It does not promise a human pressing a button before every machine action. OpenAI's 28 February 2026 defence agreement statement and Anthropic's March 2026 account of its defence dispute describe different restrictions and institutional choices. The question for a buyer is whether the limits survive software updates, subcontracting, loss of communications and emergency operating modes.
Australia's interest is therefore both civic and defensive. It needs useful automation, resilient communications and the ability to defend people. It also needs authority over how those capabilities operate locally. Buying advanced equipment while surrendering every meaningful control would leave national sovereignty resting on someone else's continuing permission.
Materials and machinery can change the rate of expansion
Compute, robotics, energy and mining are parts of a connected production system. More capable AI can help design better chips, motors and factories. More capable robots can build equipment, inspect mines and maintain facilities. Greater material and energy output can then support more computing and automation. This paper's 5 November 2026 hypothesis is that these effects can reinforce one another strongly enough to sustain periods of exponential growth. Their strength and timing have to be tested against production evidence.
Mining is one step in a longer chain. Rock must be extracted, useful material recovered, and the result refined to the quality needed by manufacturers. Components then need suitable tooling and reliable production. Copper, aluminium and steel support electrical and mechanical infrastructure. Selected batteries, magnets and electronics also require more specialised inputs. The International Energy Agency's 16 July 2026 minerals outlook projected a 25 per cent copper supply gap in 2035 against its assessed requirements and project pipeline. That is a warning about the currently identified development path, not an immutable limit on future supply.
Technology can change that path. BHP's 30 May 2023 announcement described AI-assisted recommendations for copper recovery at Escondida. Rio Tinto's 15 January 2026 agreement with Amazon Web Services linked its Nuton copper-processing technology with computing support and a data-centre customer, following first industrial deployment at Johnson Camp. These examples show ways to improve output from material already being processed and connect new extraction methods with demand. Expanded mines, refining, recycling and substitution can all contribute alongside new discoveries.
Google DeepMind's Graph Networks for Materials Exploration, or GNoME, belongs at the discovery stage. Its 29 November 2023 announcement reported 2.2 million predicted crystal structures, including about 380,000 particularly stable candidates for experimental work. The system predicts inorganic crystals and their stability. Metamaterials are a related design field in which a material's engineered structure produces useful behaviour. Both can contribute to better hardware, but a promising candidate still needs its properties measured and a workable production route.
The route from discovery to useful equipment can itself become faster. A robotic laboratory can propose an experiment, make a sample, measure it and use the result to select the next experiment. The 29 November 2023 Nature paper on the A-Lab autonomous laboratory describes this combination of computation and automated synthesis. In an Australian workshop analogy, the advance is both a much larger book of possible recipes and machinery that can test recipes, learn from the results and prepare the next batch. Scaling the successful recipe into a factory is another opportunity for engineering and automation.
There is already evidence of AI improving parts of its own industrial support system. Google DeepMind's 14 May 2025 AlphaEvolve report described deployed improvements to data-centre scheduling, chip design and model-training software. Its scheduling improvement recovered an average 0.7 per cent of Google's worldwide computing resources. This is a measured company-reported example of the feedback mechanism. More capable systems, potentially including SI, could broaden it to manufacturing methods, extraction chemistry, energy conversion and material substitution.
The hypotheses allow those improvements to change the inputs required for expansion. If a satellite design delivers twice the useful computing power for each tonne launched, the same launch fleet can deliver twice the capacity. If a motor achieves the required performance with less of a constrained material, more robots can be built from the same refined supply. These are conditional relationships assessed on 5 November 2026. No particular undisclosed material or future SI breakthrough is assigned a guaranteed arrival date.
Material demand must be checked as the hardware count grows. At an illustrative 60 kilograms per robot, annual production of 410 million robots would embody about 24.6 million tonnes of finished machines. Each one kilogram of a particular input per robot would require about 410,000 tonnes of that input annually. These 5 November calculations use illustrative quantities, not a verified Optimus bill of materials. The real test needs each design's material recipe, manufacturing losses and recycled content. Doubling production while cutting material per machine by 20 per cent still increases that material's requirement by 60 per cent.
For Australia, this creates a concrete sovereignty question. Supplying ore alone captures a different role from controlling refining, materials research, machinery design and local computing. Shared research capacity and manufacturing skills could help communities and national institutions participate in the improvement cycle. The same cycle could accelerate defence production, so control over software updates, production instructions and the allocation of resources becomes more consequential as throughput grows.
What the next decade could look like
The following five horizons are conditional hypotheses prepared on 5 November 2026. They combine dated company milestones with calculations for terrestrial computing, independently powered orbital computing and humanoid production. Exponential expansion is explicitly included. Factory, material and deployment requirements provide observable tests of each path. The scenarios are not probability estimates, and the fastest case is not an upper bound.
The terrestrial electricity reference follows the International Energy Agency's April 2026 path towards 950 terawatt-hours in 2030, then assumes 5 per cent annual growth. Orbital electricity is added separately, using the deployment hypotheses explained below. The terrestrial processing reference starts with Epoch AI's 16 September 2026 research: 20.3 million NVIDIA H100-equivalents estimated at the end of 2025, extrapolated to 31.6 million for September 2026. An H100-equivalent expresses different processors in a common performance measure, much as different engines can be compared by power. It does not measure intelligence.
The orbital ambition case follows the early deployment rates described in SpaceX's June 2026 talk, then tests a rise to 300 gigawatts of annual additions by the end of 2031. That later number is this paper's interpretation of the scale in Musk's February 2026 interview, not a dated company commitment. A smaller deployment case uses one-tenth of those rates. Both include gradual commissioning, an assumed five-year service life and average operation at 80 per cent of rated computing power. The paper holds terrestrial construction unchanged when calculating combined totals, so these totals require genuinely additional orbital equipment.
2027 is the year of delivery tests
One year ahead, the terrestrial electricity reference reaches about 635 terawatt-hours, around 2.2 times Australia's 2025 generation. Terrestrial AI processing reaches about 1.6 times the September 2026 baseline. In the orbital ambition case, gradual commissioning produces half a gigawatt of rated fleet capacity by the end of 2027 and about 1.2 terawatt-hours of orbital computing energy during the year. The combined electricity figure rounds to 636 terawatt-hours. These are 5 November 2026 calculations; the smaller orbital case contributes about 0.12 terawatt-hours.
SpaceX's October 2026 terrestrial targets make its additional computer halls a major delivery test. On its 20 October 2026 call, management also targeted first Starmind flights in 2027. The relevant orbital test is whether complete computing spacecraft arrive with working power, heat rejection and communications. Huawei's September 2026 processor roadmap and Anthropic's April supplier agreements add competing hardware milestones, while Blue Origin's TeraWave plan targets initial deployment late in 2027.
Factory milestones are now explicit in this horizon. TSMC's 15 October 2026 timetable puts its second Arizona fab into volume production in the third quarter of 2027. Samsung's 29 October update places Taylor's start in 2027. Micron's 9 July announcement targets initial wafers from its first Idaho expansion in mid-year, while SK hynix's 28 October update places its first Yongin phase in 2027. Their contribution increases as output becomes reliable and completed equipment reaches customers.
The humanoid hypotheses begin with annual worldwide production of 100,000, 500,000 or one million machines in 2027. These are this paper's 5 November starting assumptions, ranging from delayed commercial progress to a rapid opening ramp. They are anchored to the scale of the disclosed manufacturing programmes rather than a claimed census of orders. Tesla's late-2027 ramp expectation makes the million-machine starting case demanding: it requires substantial output from other manufacturers as well.
For Australia, the near-term opportunity is to build practical competence while the equipment is being installed. A community or institution can test local AI, retain its records in usable formats and practise operating through an external outage. Those are achievable steps before the largest future systems arrive.
2028 is the year manufacturing plans must become machines
Two years ahead, the terrestrial reference reaches 726 terawatt-hours, about 2.5 Australias, and roughly 2.4 times the starting terrestrial AI processing capacity. The orbital ambition case reaches six gigawatts of fleet capacity by year-end and uses about 18 terawatt-hours during 2028, bringing the combined electricity figure to about 744. The smaller orbital case brings it to about 728. These 5 November 2026 calculations allow spacecraft power to expand independently of electricity infrastructure on Earth.
The Terafab volume-production timetable announced in October 2026 points into this year. Amazon's April 2026 Globalstar agreement targets expanded direct-to-device services from 2028, while Helion's project schedule checked in November 2026 also points towards initial power delivery. Each requires a different proof: chips manufactured reliably, useful satellite coverage, or electricity delivered from a generating plant.
The conventional fab programme is also substantial: Intel's 21 October 2026 update targets Fab 62 operations by the end of 2028, and Micron's 9 July timetable places its second Idaho fab's initial wafers late in 2028. Amkor's 10 August construction update targets early-2028 completion, while GlobalFoundries' 4 November announcement places its interposer ramp in the first half. In China, Nexchip's 4 January 2026 plan targets full output from its fourth phase by year-end. A factory starting late in 2028 may make its larger contribution during 2029 and beyond.
Annual humanoid production reaches 150,000, one million or 2.5 million in the three 2028 hypotheses. The factory expansion case doubles its 2027 output; the accelerated case multiplies it by 2.5. These assumptions require improving yields, supplier output and customer use, with additional production lines opening as existing ones mature. Materials discovery contributes when a tested improvement can actually enter those manufacturing processes.
The likely pressure on buyers is integration. It becomes convenient to purchase computing, identity, communications and automated assistance together. Australia's negotiating position will be stronger where switching suppliers, moving personal records and continuing local operations are already ordinary capabilities.
2029 is the year separate systems may start to combine
Three years ahead, the terrestrial reference reaches 830 terawatt-hours, about 2.9 Australias, and roughly 3.8 times its starting AI processing capacity. The orbital ambition case reaches a 61-gigawatt fleet at year-end and about 182 terawatt-hours of orbital computing energy during 2029. Combined electricity reaches about 1,013 terawatt-hours, or 3.5 Australias. The smaller orbital case gives about 849. Louisiana's August 2026 announcement places the earliest launches from its proposed SpaceX campus in this year, while OpenAI and Broadcom's October 2025 delivery timetable runs through year-end.
GlobalFoundries' 4 November 2026 disclosure targets 2029 for its first new Malta fab to come online. The wafer-capacity index calculated from SEMI's third-quarter 2026 outlook reaches about 126, against 100 in 2026. That measures manufacturing capacity. The processing scenario measures the accumulated computing fleet.
The established industrial-robot market has a separate 24 September 2026 industry forecast of about 806,000 installations in 2029. This paper's humanoid hypotheses reach annual output of 225,000, two million or 6.25 million. The accelerated path requires rapidly repeatable factories and useful machines that justify continued production. The industrial and humanoid series are reported separately because future classifications may overlap.
Musk's February 2026 prediction puts the orbital-cost crossover by early 2029. The corresponding test is whether customers can buy useful space-based computing more cheaply over its working life. This paper's 5 November 2026 assessment leaves that outcome open because launch cost, heat rejection, useful workloads and replacement rates remain decisive. A successful orbital demonstration is an important engineering step; a competitively priced service that customers can rely on is the next test.
2031 is the year ownership choices become harder to undo
Five years ahead, the terrestrial reference reaches about 998 terawatt-hours, roughly 3.5 Australias, and about nine times the starting terrestrial AI processing capacity. In the orbital ambition case, the accumulated fleet reaches 461 gigawatts at year-end and consumes about 2,296 terawatt-hours during 2031. Combined electricity reaches about 3,294 terawatt-hours, or 11.5 Australias. The smaller orbital case gives about 1,227 terawatt-hours combined. These 5 November 2026 hypotheses show why terrestrial electricity growth cannot represent the complete system.
By then, the 2030 targets provide substantial evidence for judging the race: China's September 2026 national computing plan, the International Energy Agency's April 2026 electricity outlook and the satellite ambitions already described. Anthropic's November 2026 filing also places hundreds of billions of dollars of obligations before the end of 2030. Intel's first Ohio fab belongs near this horizon under its 27 February 2025 published operating timetable. Bringing it into the 2027 supply total would count capacity years before its last verified scheduled opening. The contractual choices being made now reach well beyond the next product launch.
Annual humanoid production reaches about half a million, eight million or 39 million in the 2031 hypotheses. The corresponding fleets built since 2027 reach about 1.3 million, 15.5 million or 64.4 million machines before the first assumed five-year retirements. The accelerated case requires successful reinvestment of AI and robotic capability into production, logistics and material processing. Evidence of robots reliably helping expand those activities would strengthen that hypothesis.
The major risk for Australian institutions is becoming dependent on a supplier for accumulated records, business procedures, staff expertise and machine control at the same time. It resembles a town discovering that its only fuel supply, freight depot and communications service all depend on one external contract. The cost of leaving can grow even when individual services are attractive. Local ownership and interoperability become much easier to establish before that dependence is entrenched.
2036 has a wide range and very large consequences
Ten years ahead, the terrestrial electricity reference reaches about 1,273 terawatt-hours, or 4.4 Australias. The orbital ambition case has 1,500 gigawatts of rated fleet capacity at year-end after five-year replacements, with about 10,395 terawatt-hours consumed during 2036. Combined electricity reaches about 11,668 terawatt-hours, equivalent to 40.7 Australias. The smaller orbital case gives about 2,313 terawatt-hours combined, or 8.1 Australias. These 5 November 2026 calculations measure computing energy supplied on Earth and in orbit. The orbital component places no equivalent demand on Earth's electrical grid.
An integrated network of orbital and terrestrial computing, robots and communications could by then perform a substantial share of industrial and administrative work. Its ownership could concentrate further or spread through public, community and personal infrastructure. The reviewed announcements do not establish which companies will dominate in 2036, a reliable US-China market split, or a guaranteed SI capability. Multi-fab campus plans make further large expansion plausible, but the publicly verified operating schedules do not extend far enough to construct a reliable ten-year factory total. The enduring issue is whether access to essential intelligence and machinery remains contestable.
Annual humanoid output in 2036 reaches about 1.5 million in the delayed case, 60.8 million with repeated factory expansion, and 409.6 million with accelerated industrial feedback. Assuming five-year service lives and deployment of the completed machines, the surviving fleets from these programmes reach about 5.2 million, 158 million and 988 million respectively. The accelerated calculation tests the industrial conditions behind Musk's September 2026 billion-humanoid expectation. It is a hypothesis about what sustained compounding could produce, with a material and manufacturing requirement that can be examined.
An optimistic outcome would make sophisticated assistance as ordinary as access to electricity or a public library, while retaining personal choice. A dangerous outcome would make participation in daily life conditional on a handful of automated systems that people cannot inspect, challenge or leave. Decisions about ownership, maintenance, legal authority and local operation will help determine which outcome becomes easier to reach.
Technologies that could change the trajectory
Quantum computing and atomically precise manufacturing deserve attention because they may change what can be built, rather than simply enlarging today's computer halls. A quantum computer uses physical quantum behaviour to perform certain kinds of calculation. It is a specialist tool with potentially important applications, rather like a laboratory instrument alongside a general workshop. A large quantum machine would not automatically replace every ordinary computer.
IBM's roadmap checked on 5 November 2026 targeted its Starling system for 2029, with 200 logical qubits and 100 million operations. A logical qubit is a quantum information unit protected by error correction, which uses additional physical resources to keep mistakes under control. IBM's later roadmap points to 2,000 logical qubits and a billion operations from 2033 onward. Those are engineering targets. Useful performance on real tasks is the relevant test. PsiQuantum's programme description checked on 5 November 2026 adds an Australian connection through plans for large systems using light-based technology.
Atomically precise manufacturing concerns placing and controlling matter at the scale of individual atoms. It can be pictured as building with a plan for each brick, taken down to the smallest structural scale. The US National Institute of Standards and Technology's programme updated in October 2026 addresses devices, measurement and manufacturability at that scale. Australia's Silicon Quantum Computing described automation of atomic-precision semiconductor manufacturing in 2026. Its Australian Government directory entry checked on 5 November 2026 records a mixture of government, university and corporate interests. These are concrete foundations for future specialised computing. They do not yet provide a public engineering basis for assuming unrestricted manufacture of arbitrary objects.
Brain-computer interfaces form a separate line of development. They connect signals from a nervous system with external equipment. Neuralink's January 2026 update reported 21 trial participants. Such work may create valuable assistive options. For a sovereignty framework, adoption remains a personal choice. Access to ordinary services, personal AI or civic participation should not depend on accepting an implant.
Unidentified anomalous phenomena, often shortened to UAP, and claims of concealed or non-human technology do not supply a reproducible engineering baseline in the sources reviewed for this paper. They are therefore not assigned computing capacity or deployment dates in the projections. New evidence could change that assessment. The scale established through disclosed spending, equipment and construction is already sufficient to justify serious preparation.
The missing middle for sovereignty
The preceding sections describe systems large enough to shape national choices. Luke Nathan Hayes's proposal, A Fair Go in the Age of Superintelligence, asks what physical and civic capacity people need between their own devices and those vast systems. Its public submission and discussion of possible futures, checked on 5 November 2026, place that question within a broader argument for shared capability and self-determination.
The missing middle in A Fair Go in the Age of Superintelligence is shared local computing infrastructure. Hayes's supplied compute submission, dated 18 August 2026, proposes racks that give communities serious processing capacity they can govern and use. A household computer serves one scale of need; a distant industrial data centre serves another. Locally controlled equipment supplies useful capacity between them, while connecting with larger systems when people choose.
Think of the role of a public library, a workshop at a technical and further education college, or TAFE, and a local emergency-services facility. Each gives people access to equipment, skills or resilience that most households could not efficiently provide alone. A computing facility could offer related benefits: machines, training, practical assistance, secure local storage and continuity during an outside disruption. The exact ownership arrangement would have to be chosen and documented locally.
The physical proposal requires more than a cabinet of expensive chips. Electricity, cooling, communications, trained maintainers, replacement budgets and rules for allocating access all belong in the operating plan. A locally owned machine can still depend on foreign chips or software. Sovereignty grows when those dependencies are understood, several workable options exist and an external service failure does not disable every essential function.
Hayes's 18 August 2026 submission used Australian 6 million dollars per location as an initial planning allowance covering equipment and supporting facilities. At that assumed cost, 100 locations would require Australian 600 million dollars, and 1,000 would require Australian 6 billion. These calculations, made on 5 November 2026 from the submission's allowance, show the order of funding involved. The original worldwide thought experiment applied the same allowance to 44,000 locations, reaching Australian 264 billion dollars. The next useful step is a costed pilot with measured use and operating expenses.
The supplied submission names Dunwich on Minjerribah, also known as North Stradbroke Island, as a possible starting location. That local starting point does not limit the purpose to Redlands. The infrastructure idea concerns personal sovereignty, local data sovereignty, national resilience and self defence. Practical self defence includes protecting records, resisting unauthorised control, maintaining communications and retaining the capacity to operate through disruption.
Individual Auras use that capacity
Aura of Intelligence is the software project at the centre of Hayes's work. The intended arrangement is for people to have individual Auras: personally controlled AI that can learn their authorised context and help them use available tools and computing resources. The Aura of Intelligence public site and Personal Aura page, checked on 5 November 2026, provide entry points. Hayes's design calls for open-source software that people can inspect and modify, personally owned models, and agents that carry out authorised tasks. It has no required dependence on OpenAI or another frontier-model provider. Licences and access to the model files must allow those ownership promises to work in practice.
An individual Aura is closer to a person's chosen assistant and working memory than to the local computer facility itself. It could help organise a project, find relevant information, prepare a document or use additional local processing when a task exceeds the person's own device. The individual remains in charge of what it may remember, which tools it may use and when it should stop.
Keeping information, retrieving it and changing a model are separate operations. A personal archive can hold documents and memories. Search can bring selected material into a current task. Further training can change how the model responds. A meaningful ownership arrangement must make these choices visible rather than silently mixing them together. It also needs practical export, backup and migration so the person can move their records and working setup.
A digital twin is a digital representation used to describe or simulate something in the world. A farm twin might represent paddocks, rainfall and equipment; a personal representation might hold someone's preferences and chosen history. Its usefulness depends on accuracy, correction and consent. The person represented must be able to reject an inference, change a record or stop participating. Hayes's supplied whole-of-life Aura material places personal continuity and practical support within this software ambition.
Consider an illustrative community project, constructed for this paper on 5 November 2026. A resident on Minjerribah wants to help a local food garden plan water use. Their Aura helps assemble the questions and keeps their private calendar and personal notes on their chosen equipment. A community-owned rack performs a heavier simulation using permitted weather, soil and infrastructure data. The rack provides the machinery. The individual Aura helps its owner use it.
The Web3 sensorium connects what people choose to share
The Web3 sensorium is the proposed intermediary layer between individual Auras, devices, shared computing and wider information. Sensorium refers to the ways a system gathers and makes sense of observations. Web3, in this proposal, concerns distributed control and verifiable cooperation between independently operated systems. Hayes's supplied c-hour plan, dated 23 September 2026, explicitly describes participation without requiring cryptocurrency purchases or expertise.
Think of it as a combination of a switchboard and a map room. It helps the right information reach the right place, while preserving who supplied it, when it was measured and what permissions apply. The public Web3 Sensorium page, dated 9 May 2026 and checked on 5 November, describes a research direction involving local control, shared models, evidence and public debate. It remains a proposed architecture whose protections have to be implemented and tested.
In the garden example, the sensorium could connect an authorised soil reading with a public weather forecast and the shared water model. It would not need the resident's private messages or complete personal history. If a result were shared publicly, people should be able to inspect the source, date and assumptions. The Minjerribah Living Twin public project, checked on 5 November 2026, provides a related place-based entry point for exploring digital representations.
This intermediary layer has an important trust problem to solve. Copies already disclosed cannot always be recalled by pressing a revoke button. Consent therefore needs to govern collection and disclosure in advance, with limited retention, clear recipients and enforceable obligations. A promise of personal control becomes credible through demonstrated behaviour: what leaves the device, who can decrypt it, what happens offline and whether the person can leave with usable records.
Purple Party for Australia provides the proposed civic layer
P4A stands for Purple Party for Australia. Its public civic workbench, checked on 5 November 2026, describes a proposed movement and drafting project. In Hayes's architecture, it aims to provide the civic layer: a place for people to understand the rules that govern them, examine evidence, develop changes and build support through democratic processes.
Its scope extends beyond one council. A resident may first encounter a problem through a local road, housing issue or service, while the power to change it sits with a state parliament or the Commonwealth. P4A's intended role is to help people follow that chain, understand what a proposal would do and work towards an informed, winnable pathway to change. Other countries would need processes fitted to their own institutions.
An Australian analogy is a community meeting that has access to a good library, clear maps and time to rehearse the consequences of a proposal. People still disagree and decide. The tools help them see which authority is responsible, which claims have evidence and what trade-offs they are being asked to accept. The Australian Law and Personal Relevance public project, checked on 5 November 2026, is a related entry point into that learning work.
A constitutional referendum has a particular threshold. The Australian Electoral Commission's explanation checked on 5 November 2026 states that a change requires a national majority of voters and a majority in at least four of the six states. A simulation or local vote does not enact the change. P4A's proposed contribution is the groundwork: understanding the law, testing possible wording, listening to objections and building the public support needed for the formal process.
Returning to the garden example, the sensorium might show that an apparent water problem also involves a planning rule or public infrastructure decision. Individual Auras could help participants understand the material. The shared rack could run agreed scenarios. P4A could support the civic discussion about a proposed change. The participants, through the appropriate democratic institutions, would retain the decision. This is how the physical, personal, intermediary and civic parts fit together.
A reason to keep participating
Infrastructure becomes socially useful when people can build worthwhile lives around it. Hayes's c-hour, or community-hour, proposal addresses reciprocity: recognising contributions to other people and the community without assigning those hours a monetary equivalent. The C Hours introduction, checked on 5 November 2026, sits alongside the supplied 23 September 2026 planning document. Individuals and communities decide what they are willing to offer to people earning c-hours. In the wider braided-economy proposal, monetary activity and voluntary reciprocity can coexist without requiring every contribution to become a price.
The garden provides a simple illustration. Someone contributes time to watering or teaching. A participating group might choose to offer access to a workshop, an outing or another activity. The offer is voluntary and locally decided. The hour is a record of contribution, with no fixed conversion into dollars. Whether the arrangement builds trust depends on fair access, sensible verification and the freedom not to participate.
The Global Association for Joyful Responsible Abundance on Earth, shortened to GAJRA Earth, is Hayes's proposed voluntary association network. Its public project page, identified in the supplied material and checked on 5 November 2026, provides a starting point. Joy asks whether life is worth participating in. Responsibility asks what the activity does to other people and the living world. Abundance asks whether capability can expand access and possibility. Keeping all three together provides a practical test of what the technology is for.
Culture belongs in that account. Hayes's A Protopian Gambit music project, checked on 5 November 2026, expresses a wager on building a future people want to keep improving. Shared worlds, music, humour and art can give technological participation a human purpose. A decentralised network, with independently controlled worlds connected by choice, would be judged by whether people can create, cooperate, disagree and leave while keeping meaningful control of their own lives.
That right to leave is central to the whole architecture. People should be able to disconnect, decline a model's advice, refuse an implant and keep private parts of life outside a shared system. An individual Aura, a local rack, a sensorium connection or a civic organisation earns trust by making those choices workable.
What would make sovereignty real
The immediate task is to turn the proposed local capacity into something people can examine and use. A pilot should demonstrate useful work, an affordable operating model and continuity through an outside outage. It should show a person moving their Aura and records, a community deciding who administers the equipment, and an independent reviewer checking what privileged operators can do. These are recommendations from this paper, dated 5 November 2026.
Ownership must include practical authority. A community facility needs clear rules for selecting maintainers, approving sensitive changes, resolving disputes and paying for replacement equipment. No single administrator should silently acquire control over every personal record or connected machine. The ability to pause a service locally, preserve evidence and restore a known working version is as important as headline processing speed.
There are existing public-capacity programmes to learn from. The US National Science Foundation's 2026 account of its National Artificial Intelligence Research Resource reported support for hundreds of teams and thousands of students. The European Commission's AI factories programme checked on 5 November 2026 links shared computing with assistance for users. The UK Government's 2025 Compute Roadmap committed public investment towards 2030. These programmes have different purposes and governance arrangements, but show that access to substantial computing can be organised beyond a direct personal subscription to a single commercial model provider.
Australia can participate in the global build-out while developing its own capacity to choose. The major corporations supply equipment, research and services of enormous potential value. Their military links, concentrated voting power and control over critical dependencies make scrutiny essential. Shared local racks, individually controlled software, carefully governed connections and informed civic participation address different parts of that problem.
The central choice is whether people gain capability alongside the machines. The industrial scale is already immense, and its consequences reach well beyond a better chatbot. Building useful, accountable capacity close to people is a practical way to make personal sovereignty, local resilience and national self defence part of the same technological future.
How the projections were calculated
This section records the assumptions and calculations prepared on 5 November 2026. It separates terrestrial electricity, orbital power, comparable processor capacity and robot production. Company targets provide dated reference points; interpolations, growth rates and service lives are this paper's assumptions. The scenarios test possible expansion mechanisms without assigning probabilities. Mining, refining, materials innovation and AI-assisted improvement can change the growth rates. They are treated as conditions to measure, rather than permanent limits or automatic breakthroughs.
Terrestrial electricity
The terrestrial reference starts at 485 terawatt-hours in 2025 from the International Energy Agency's 16 April 2026 report. From 2025 to 2030, the low case grows by 10 per cent annually, the reference case by about 14.392 per cent and the high case by 20 per cent. The reference reaches 950 terawatt-hours in 2030. From 2030 to 2036, the assumed rates are 3, 5 and 12 per cent respectively. This includes power supplied to terrestrial data centres through grids or local generation. It excludes the additional orbital fleet modelled below.
For 2027, the low, central and high electricity results are 587, 635 and 698 terawatt-hours. For 2028, they are 646, 726 and 838. For 2029, they are 710, 830 and 1,006. For 2031, they are 805, 998 and 1,352. For 2036, they are 933, 1,273 and 2,382. Results are rounded to whole terawatt-hours.
Every Australian comparison divides the relevant result by 286.8 terawatt-hours, Australia's calendar-2025 generation in the 2026 government release. Continuous-power examples multiply gigawatts by 8,760 hours and divide by 1,000 to obtain terawatt-hours in a 365-day year. They assume continuous use at the stated power. Electricity generation, equipment rating and delivered computing work remain distinct quantities.
Comparable terrestrial AI processing capacity
The starting point is 31.6 million H100-equivalents from Epoch AI's 16 September 2026 research. Epoch estimated 20.3 million at the end of 2025, then extrapolated to September 2026; the latter was not a direct September census. H100 is the reference NVIDIA accelerator used to compare different equipment. Epoch's methodology checked on 5 November 2026 explains how equipment is compared and the allowance for deployment delays. This paper's model estimates theoretical processing capacity. Useful output also depends on memory, communications, software and the work being performed.
From the 2026 baseline to 2031, the terrestrial low case assumes powered capacity grows 10 per cent annually and processing per unit of power improves 15 per cent. The reference assumptions are 20 and 30 per cent; the high assumptions are 25 and 40 per cent. Multiplying these factors gives annual processing multipliers of 1.265, 1.56 and 1.75. These are assumptions made on 5 November 2026, informed by scaling research including Epoch AI and the Electric Power Research Institute's 11 August 2025 study. They already allow compounded growth; they are not a complete model of what future SI could improve.
From 2031 to 2036, the illustrative terrestrial cases use powered-capacity growth of 3, 8 and 15 per cent annually, with hardware efficiency gains of 10, 20 and 30 per cent. The resulting multipliers are 1.133, 1.296 and 1.495. This selected slowdown is a scenario assumption, not a claim that continued faster expansion is impossible. AI-assisted chip design, new processes and greater production could justify a different path. Software improvements to useful work per calculation should be reported separately from physical processor capacity.
The terrestrial processing and electricity series remain separate sensitivity calculations for different equipment populations. They are not a single solved energy balance. Combining them into one physical fleet forecast would require measured processor performance per watt, the equipment mix and utilisation. The new orbital calculations address power and deployment explicitly without inventing that missing processor conversion.
The terrestrial results, in millions of H100-equivalents, are 40, 49 and 55 for 2027; 51, 77 and 97 for 2028; 64, 120 and 169 for 2029; 102, 292 and 519 for 2031; and 191, 1,067 and 3,873 for 2036. Each sequence gives low, reference and high. The orbital power cases are not converted to H100-equivalents because their future processor mix and measured performance are not established. A combined processor total would need those measurements and must count each physical machine only once.
How factory capacity enters the scenarios
The factory review checks major publicly disclosed projects against the five horizons. It includes advanced processor fabs, memory and storage plants, packaging, and factories making supporting electronics. Semiconductor companies' investment totals are not converted directly into computing capacity. A new building may still need equipment; installed equipment may still need qualification; initial output can take time to reach the planned rate.
To calculate processing capacity from the factories, start with wafers processed each month, allow for the production ramp, and count the usable chips per wafer. Then establish how many are allocated to AI products. Matching memory and packaging must be available before completed systems can be shipped, connected and powered. The running fleet accumulates those additions and loses equipment through retirement. This resembles estimating the grain available in a town: the silo's handling capacity, this year's deliveries and the stock already stored are three different figures.
SEMI's third-quarter 2026 public summary supports multiplying 100 by 1.08 annually after 2026, through 2029 only. This review used the public summary, not the full paid database. Company checks cover major disclosed projects rather than every plant. Faster processing growth additionally requires better chips, improved production yields, more output allocated to AI and sustained deployment. Wafer-capacity growth alone cannot validate the central processing scenario.
The calculations avoid counting the same capacity twice. The named projects sit within the broader industry expansion, so their totals are not added on top of SEMI's aggregate. Plants already producing before the baseline contribute only subsequent growth. An acquisition changes ownership without itself creating more worldwide production. Processor, memory and packaging capacity describe complementary stages and cannot be summed as equivalent processors. Missing Chinese production details remain an explicit uncertainty.
The International Energy Agency's 16 April 2026 report already considered semiconductor and memory supply constraints in its energy outlook. That does not establish that its April projection included each later factory announcement. The new construction evidence makes the industrial pipeline more visible, especially in 2027 to 2029, without resolving undisclosed AI allocations, usable output, retirement rates or the power available to operate the resulting equipment. The retained numerical cases should therefore be read as conditional scenarios, not forecasts validated by every fab in construction.
Orbital power and accumulated capacity
The orbital ambition case uses year-end annual commissioning rates of one gigawatt in 2027, ten in 2028 and one hundred in 2029, following the scale in the 8 June 2026 SpaceX technical talk. This paper selects 200 in 2030 and 300 in 2031, then holds the annual rate at 300 through 2036. The 300-gigawatt choice represents one numerical interpretation of the several-hundred-gigawatt scale in Musk's 5 February 2026 interview. The intermediate years and the continuation after 2031 are assumptions dated 5 November 2026, rather than company guidance.
Commissioning rates increase smoothly between those year-end points. The incremental fleet starts from zero at the end of 2026; that modelling convention excludes existing demonstrations rather than claiming they do not exist. For example, a rate rising from zero to one gigawatt per year over 2027 delivers half a gigawatt during that year. A rate rising from one to ten during 2028 delivers another 5.5 gigawatts, taking the end-year fleet to six. An annual deployment rate is therefore kept distinct from the installed fleet.
Each commissioned unit is assumed to carry adequate generation and heat-rejection equipment for its rated computing load. The model assumes average operation at 80 per cent of that rating and replacement after five years. Both are this paper's 5 November assumptions, not measurements of a future fleet. The average operating fraction covers the difference between rated and realised electrical load, including outages and derating. It does not apply an Earth-grid availability constraint. Replacement removes old equipment from the stock before the next year's total is calculated.
Energy equals the average surviving rated fleet during the year, multiplied by 80 per cent and 8,760 hours. Converting gigawatt-hours to terawatt-hours requires dividing by 1,000. The ambition case's orbital energy is approximately 1.2 terawatt-hours in 2027, 17.5 in 2028, 182 in 2029, 2,296 in 2031 and 10,395 in 2036. The smaller case uses one-tenth of every commissioning rate and therefore one-tenth of these energy figures. Both scenarios are conditional on successful manufacturing and deployment; a multi-year delay would shift the trajectory.
The reported combined totals add orbital consumption to the unrounded terrestrial reference. They describe electrical energy used by the two computing systems, not new demand on terrestrial grids. Orbital power refers to the computing load; spacecraft support equipment, manufacturing, mining, robot charging and launch energy are outside that comparison. Adding the two paths assumes extra orbital production and demand alongside terrestrial construction. If the same chips or customer workloads move from land to orbit, the terrestrial path must be reduced accordingly. The chip supply cannot be counted twice.
Using the earlier assumption of 100 kilowatts of rated compute per tonne, 100 gigawatts of annual additions requires approximately one million tonnes of complete computing payloads delivered annually. At 300 gigawatts, it requires three million tonnes. With 100 tonnes per dedicated flight, those are 10,000 and 30,000 successful computing-payload flights respectively. These are 5 November 2026 requirements of the hypothesis, not a forecast that all launches carry computers. A verified doubling of usable compute per delivered tonne would halve the mass and flight requirements.
Operating fraction and service life have large effects. For a fleet with the same deployment history, 100 per cent average operation would produce 25 per cent more computing energy than the 80 per cent assumption. Once a constant 300-gigawatt annual deployment programme matures, three-year replacements sustain about 900 gigawatts of stock and five-year replacements sustain about 1,500. These sensitivities identify useful engineering measurements instead of treating the Sun's energy supply as the limiting quantity.
Humanoid production and useful robot fleets
The established industrial-robot reference remains the International Federation of Robotics' 24 September 2026 forecast: 655,000 annual installations in 2026 and 806,000 in 2029. A constant-growth interpolation gives about 702,000 in 2027 and 752,000 in 2028. These figures cover that established market. The new humanoid hypotheses below address a separate emerging production system and do not use the industrial forecast as a ceiling.
The delayed adoption hypothesis assumes worldwide humanoid production of 100,000 in 2027, growing 50 per cent annually through 2031 and 25 per cent thereafter. The factory expansion hypothesis starts at 500,000, doubles annually through 2031 and then grows 50 per cent annually through 2036. The accelerated hypothesis starts at one million, multiplies production by 2.5 annually through 2031 and then by 1.6 annually through 2036. All starting values and multipliers are assumptions selected on 5 November 2026. They represent increasingly successful commercialisation and industrial improvement, rather than measured orders or confidence intervals.
The starting assumptions span partial realisation through rapid realisation of the disclosed manufacturing programmes. The first million-unit Tesla line is a useful scale reference, with several competing manufacturers contributing globally. Continued growth requires replication of productive lines, expansion of component supply and useful deployments that support reinvestment. The later growth reductions are selected scenarios, not mandatory laws. Faster, slower or interrupted paths remain possible. A single early factory's rapid improvement does not establish the same growth rate for every supplier indefinitely.
Annual production is calculated by repeatedly multiplying the previous year's output by the applicable growth factor. The five forecast years give 100,000, 150,000, 225,000, 506,250 and about 1.54 million in the delayed case. The factory expansion case gives 500,000, one million, two million, eight million and 60.75 million. The accelerated case gives one million, 2.5 million, 6.25 million, 39.06 million and 409.6 million. These calculations explicitly include exponential growth over each period.
For fleet calculations, completed robots are assumed to enter service during the production year and remain for five years. At the end of 2036, the model therefore retains production from 2032 through 2036, giving approximately 5.2 million, 158.3 million and 988.1 million robots. Older machines are retired. These are surviving units under the assumptions, not human-worker equivalents. Useful labour depends on availability, productive hours, task capability and the amount of human support required. Those measures also determine how much robot deployment can accelerate mining and manufacturing.
Testing the feedback hypothesis
The production paths must be checked against their changing physical requirements. Annual material demand equals units produced multiplied by material per unit, adjusted for manufacturing losses and recycled supply. For orbital systems, useful power per tonne links the capacity scenario to spacecraft mass and launches. For robots, accepted output per line and the number of productive lines link annual production to factories. These relationships make the growth hypotheses testable.
The main indicators are completed output, yield, machine availability, useful work per hour, the time required to open another production line, refined material delivered and material required per unit. For AI-assisted discovery, the relevant sequence is predicted candidate, reproducible measurement, qualified component and manufacturing-scale adoption. Improvements at any stage can accelerate the loop; a sustained shortfall changes the achievable rate. Shared chip and material supplies must be allocated across programmes before combined scenarios can be treated as a consistent world forecast.
The stronger hypothesis gains support if AI and robots measurably shorten factory expansion, improve recovery and refining, reduce material requirements or increase useful computing per watt and per tonne. It weakens if those improvements fail to reach production while costs, failures or delivery delays rise. Uncertainty about this feedback does not make exponential expansion impossible. It makes the feedback strength and the time needed to realise it central research questions. No extra unspecified SI multiplier is added on top of the already compounded numerical cases.
Reading the evidence
Financial filings provide dated disclosures of spending, commitments and voting rights. Company announcements provide reported operations and management targets. Government publications establish the permissions, plans or figures they describe. Independent studies bring their own coverage and estimation methods. A filing's publication date and the period measured are both retained where available. Where a public page provides no reliable publication day, this paper gives the available month or year and the date checked.
The local sovereignty sections draw on Hayes's supplied documents and the public project pages linked in the text. The supplied compute submission is dated 18 August 2026; the supplied c-hour planning edition is dated 23 September 2026. The individual Aura and layer definitions also incorporate Hayes's explicit clarifications in this discussion. Proposed capabilities are described as design intentions. An accessible public page demonstrates published material, rather than independent proof that every described system is deployed.
Reference register
The links in the main text open the relevant source directly. This register supplies publication details in order of first use. Sources with no reliable publication day retain the month, year or access date available. The public project pages appear in the sovereignty discussion rather than as unexplained repository links.
1. SpaceXSI. 22 October 2026. Introducing the SpaceX Super Intelligence Partner Network.
2. Walter Isaacson interviewing Elon Musk for The Builders. Recorded 6 August 2026; published 20 October 2026. Interview on Super Intelligence control work and abundance. Original video. Rehosted interview transcript.
3. Australian Government Department of Climate Change Energy the Environment and Water. 2026 release; checked 5 November 2026. Australian Energy Statistics Table O electricity generation in 2024 to 2025 and calendar 2025.
4. IEA. 16 April 2026. Key Questions on Energy and AI: executive summary.
5. SpaceX. 20 October 2026. Third-quarter 2026 earnings release.
6. SpaceX / SEC. 22 October 2026. Quarterly report, September 2026.
7. SpaceX management. 20 October 2026. Third-quarter earnings call transcript, rehosted by The Motley Fool.
8. Tesla / SEC. 21 October 2026. Third-quarter 2026 update, exhibit 99.1.
9. SpaceX / SEC. 11 June 2026. Final IPO prospectus.
10. Starlink. 6 October 2026 update. Network updates.
11. NVIDIA. 24 August 2026. SpaceXAI adopts NVIDIA Vera CPU.
12. FCC. 4 February 2026. SpaceX orbital data-center application accepted for filing, DA 26-113.
13. Louisiana Economic Development. 25 August 2026. SpaceX Louisiana campus announcement.
14. FAA. Notice of intent released 14 October 2026; page checked 5 November 2026. SpaceX Starship environmental review materials.
15. Google Research. 4 November 2025. Project Suncatcher system design and integrated solar power.
16. SpaceX technical talk with Elon Musk and Ian Dahl. 8 June 2026; transcript checked 5 November 2026. AI satellites at scale and annual deployment targets. Original video.
17. Dwarkesh Patel and John Collison interviewing Elon Musk. Published 5 February 2026. Interview on orbital computing manufacturing and AI.
18. Amazon. Checked 5 November 2026. What is Amazon Leo / Project Kuiper?.
19. Amazon. 14 April 2026. Agreement to acquire Globalstar and expand Amazon Leo.
20. Blue Origin. 2026; checked 5 November. Introducing TeraWave.
21. Google Research. 30 October 2026. Project Suncatcher's first satellite AI response.
22. State Council Information Office. 14 February 2026. China advances space computing tests.
23. Beijing International Science and Technology Innovation Center. 28 January 2026. Orbital computing constellation plans.
24. OpenAI. Checked 5 November 2026. Our structure.
25. OpenAI. 31 March 2026. Accelerating the next phase of AI.
26. OpenAI. 29 April 2026. Building the compute infrastructure for the Intelligence Age.
27. OpenAI. 21 January 2025. Announcing the Stargate Project.
28. OpenAI / Broadcom. 13 October 2025. Strategic collaboration to deploy 10 gigawatts of AI accelerators.
29. OpenAI / NVIDIA. 22 September 2025. Strategic partnership announcement.
30. OpenAI / Amazon. 27 February 2026. Strategic partnership.
31. Reuters. 13 May 2026. Altman has over $2 billion stake in companies that dealt with OpenAI.
32. Tools for Humanity. Checked 5 November 2026. Company website.
33. Oklo. 22 April 2025. Chairman transition.
34. Helion. Checked 5 November 2026. Orion power plant.
35. Oklo / Meta. January 2026. Agreement supporting 1.2 GW nuclear development in Ohio.
36. Anthropic / SEC. 2 November 2026. Amended registration statement, S-1/A.
37. Anthropic. 28 May 2026. Series H financing.
38. Anthropic. 20 April 2026. Expanding our use of Amazon compute.
39. Anthropic. April 2026. Google and Broadcom compute partnership.
40. Alphabet. October 2026. Third-quarter 2026 earnings call.
41. Alphabet / SEC. 2026; ownership as of 6 April. Prospectus supplement: voting control disclosure.
42. Amazon. 29 October 2026. Third-quarter 2026 results.
43. Amazon / SEC. October 2026. Quarterly report, September 2026.
44. Reuters, via Investing.com. 17 November 2025. Bezos to co-lead AI startup in first operational role since Amazon.
45. Microsoft. 28 October 2026. FY2027 first-quarter earnings.
46. Meta. 21 October 2026. Third-quarter 2026 results.
47. Meta / SEC. 2 November 2026. Special-meeting proxy: proposed non-voting Class C shares.
48. NVIDIA / SEC. 12 May 2026. 2026 proxy statement.
49. NVIDIA. Checked 5 November 2026. NVentures portfolio.
50. NVIDIA. 2026; checked 5 November. AI infrastructure financing platforms to mobilise over $500 billion.
51. SEMI, distributed by PR Newswire. 13 October 2026. Global fab equipment spending and construction pipeline.
52. SEMI. Third quarter 2026 edition; checked 5 November 2026. 300 millimetre Fab Outlook public highlights and methodology.
53. TSMC. 15 October 2026. Third quarter 2026 earnings call transcript.
54. Arizona Commerce Authority. 16 July 2026. TSMC additional Arizona investment.
55. Intel. 9 October 2025. Fab 52 and Intel 18A production announcement.
56. Intel Foundry, republished by IMAPS 3D InCites. 21 October 2026. Building Americas capacity to innovate.
57. Intel, letter to Ohio Department of Development, hosted by Engineering News Record. 27 February 2025; construction update cross-checked 21 October 2026. Ohio onshoring report and revised construction timetable.
58. Samsung Electronics. 29 October 2026. Third quarter 2026 results and Taylor outlook.
59. Micron. 9 July 2026. Accelerated US investment and first concrete at New York fab.
60. SK hynix. 28 October 2026. Third quarter 2026 results and memory capacity expansion.
61. GlobalFoundries. 4 November 2026. Third quarter 2026 results and manufacturing expansion.
62. Arizona Commerce Authority. 10 August 2026. Amkor packaging construction update.
63. Texas Instruments. 24 February 2026. Capital management presentation.
64. ASML / SEC. 14 October 2026. Third quarter 2026 results and equipment capacity outlook.
65. State Council of China. 7 September 2026. Information and communication infrastructure plan.
66. State Council Information Office. 26 May 2026. Press briefing on computing infrastructure.
67. Huawei. September 2026. Huawei Connect 2026 keynote.
68. Huawei. September 2026. Ascend 960 supernode.
69. Huawei. Checked 5 November 2026. Company facts.
70. Alibaba. February 2025. AI and cloud infrastructure investment announcement.
71. Alibaba. September 2026. Apsara 2026 infrastructure roadmap.
72. Alibaba / SEC. 2026. June-quarter 2026 results.
73. Tencent. 18 March 2026. 2025 annual results presentation.
74. Epoch AI. 27 October 2026. Introducing the AI Chip Imports Explorer.
75. China Media Group and CCTV. Recorded 22 September 2026; published 23 September 2026. Musk on Chinese chipmaking and humanoid robot growth.
76. SMIC / Hong Kong Exchanges. 13 August 2026; measurements at June 2026. Second quarter 2026 results and capacity.
77. Hua Hong Grace / Hong Kong Exchanges. Published 2 September 2026; measurements at June 2026. Interim report and Wuxi Phase Two equipment installation.
78. Hua Hong Grace / Hong Kong Exchanges. 1 September 2026. Wuxi Phase Three joint venture investment agreement.
79. CXMT / CNINFO. Published 29 August 2026; balances at 30 June 2026. Interim report including construction in progress.
80. China Electronics procurement platform. 10 October 2026. Shanghai CXMT production line construction procurement.
81. CXMT. 20 September 2026. Fifth generation memory process enters mass production.
82. YMTC procurement through Hubei Complete Tendering, mirrored by Zhiliaobiaoxun. 5 February 2026. Wuhan Fab Three network installation tender.
83. Hubei Complete Tendering on behalf of YMTC. 13 March 2026. Wuhan network installation preferred bidder notice.
84. China Electronics procurement, mirrored by Qianlima. 4 November 2026; mirror checked 5 November 2026. YMTC Fab Three detector procurement notice.
85. Nexchip. 4 January 2026; company page checked 5 November 2026. Hefei Phase Four construction and capacity plan.
86. China MIIT. September 2026. Qianfan constellation progress.
87. GSMA. 2026. Satellite connectivity, MWC Shanghai 2026.
88. International Federation of Robotics. 24 September 2026. Five million robots now operate in factories globally.
89. Tesla. 22 April 2026. First quarter 2026 update and Optimus factory design capacities.
90. Tesla management, transcript published by Benzinga. 21 October 2026. Third quarter 2026 earnings call on Optimus production.
91. Figure. 29 April 2026. Ramping Figure 03 production.
92. Figure. 15 March 2025. BotQ high-volume manufacturing facility.
93. Hyundai Motor Group. 5 January 2026. AI robotics manufacturing and Atlas deployment strategy.
94. Unitree Robotics. Statement dated 22 January 2026; checked 5 November 2026. Clarification of 2025 humanoid production and deliveries.
95. AGIBOT. 24 September 2026. Twenty thousandth robot and Chimelong deployment.
96. Reuters. 13 May 2026. Anduril raises $5 billion, valued at $61 billion.
97. US Army. 13 March 2026. Enterprise contract for IT commercial solutions.
98. JobsOhio. 27 July 2026. First Ohio built Fury rolls off the Arsenal 1 production line.
99. JobsOhio. 9 July 2025. Anduril investment and development grant agreement.
100. Ohio Tax Credit Authority. 27 January 2025. Meeting minutes and Anduril job creation tax credit approval.
101. US Army. 31 July 2025. Enterprise service agreement.
102. US Army. 22 June 2026. Next-generation command-and-control common data baseline.
103. US Army. 1 September 2026. TITAN moves to production.
104. US Government Accountability Office. 30 June 2025; describes April 2025 awards. National Security Space Launch and cost recovery report GAO 25 107228.
105. Congressional Research Service. Updated 3 September 2026; April 2026 budget request. FY2027 Defense Budget funding for selected weapon systems R49023.
106. CDAO. 14 July 2025. Partnerships with frontier AI companies.
107. US Defense Department, reprinted by GlobalSecurity. 1 May 2026. AI partnerships for classified networks.
108. Amazon / US Army C5ISR. 13 October 2026. Satellite network demonstration.
109. MITRE. 2 October 2026. Project Meridian.
110. L3Harris / Shield AI. 16 July 2026. Autonomous electronic warfare flight test.
111. Shield AI / Kraken Technology. 28 September 2026. Autonomous maritime teaming demonstration.
112. Australian Defence Ministers. 10 September 2025. Royal Australian Navy Ghost Shark acquisition.
113. Australian Defence Ministers. 19 October 2026. Sydney press conference on Ghost Shark payloads.
114. Australian Defence Ministers. 29 January 2026. Counter drone services panel for Defence and other government agencies.
115. AusTender notice mirrored by Insight Portal. Published 11 February 2026; amended 8 April 2026; mirror checked 5 November 2026. CN4218984 Defence information system platform. Official AusTender notice.
116. AusTender notice mirrored by Insight Portal. Published 29 October 2024; mirror checked 5 November 2026. CN4104368 Defence data services. Official AusTender notice.
117. AusTender notice mirrored by Insight Portal. Published 30 January 2023; latest listed amendment 9 January 2026; mirror checked 5 November 2026. CN3942923 AUSTRAC data analytics capability. Official AusTender notice.
118. AusTender notice mirrored by Insight Portal. Published 10 April 2026; mirror checked 5 November 2026. CN4231780 AUSTRAC transaction reporting and integration platform. Official AusTender notice.
119. AusTender notice mirrored by Insight Portal. Published 18 February 2026; mirror checked 5 November 2026. CN4220255 AUSTRAC transaction reporting implementation and support. Official AusTender notice.
120. Buying for Victoria. Contract 13 December 2024 to 31 March 2028; checked 5 November 2026. PABDJR116076 Support of Centurion Platform.
121. AusTender notice mirrored by Insight Portal. Published 20 August 2023; amendment listed in August 2024; mirror checked 5 November 2026. CN3996124 Australian Criminal Intelligence Commission software support. Official AusTender notice.
122. Australian Broadcasting Corporation. 30 September 2026; holdings at June 2026. Future Fund portfolio including weapons makers SpaceX and Palantir.
123. All-In interviewing Elon Musk and Gwynne Shotwell. Published 15 September 2026; recording date not separately established. AI risks peer review Starship and Terafab. Rehosted interview transcript.
124. Microsoft. 20 July 2024. Helping customers through the CrowdStrike outage.
125. NSA / CISA. June 2023. Defending continuous integration and delivery environments.
126. US Department of Defense. 25 January 2023. Directive 3000.09: Autonomy in Weapon Systems.
127. OpenAI. 28 February 2026. Agreement with the Department of War.
128. Anthropic. March 2026. Where things stand with the Department of War.
129. International Energy Agency. 16 July 2026. Global Critical Minerals Outlook 2026 executive summary.
130. BHP. 30 May 2023. AI assisted copper recovery at Escondida.
131. Rio Tinto. 15 January 2026. Nuton copper processing and Amazon Web Services collaboration.
132. Google DeepMind. 29 November 2023. GNoME crystal predictions and materials discovery.
133. Szymanski and colleagues, Nature. 29 November 2023. An autonomous laboratory for the accelerated synthesis of inorganic materials.
134. Google DeepMind. 14 May 2025. AlphaEvolve algorithm discovery and deployed computing improvements.
135. Epoch AI. 16 September 2026. Scaling AI data centers research.
136. IBM. Checked 5 November 2026. Quantum roadmap.
137. PsiQuantum. Checked 5 November 2026. About the company and utility-scale systems.
138. NIST. Updated October 2026. Atom-scale devices: engineering, metrology and manufacturability.
139. Silicon Quantum Computing. 2026; checked 5 November. Automating atomic-precision semiconductor manufacturing.
140. Australian Government Directory. Checked 5 November 2026. Silicon Quantum Computing Pty Ltd.
141. Neuralink. January 2026; checked 5 November. Company updates.
142. Australian Electoral Commission. Checked 5 November 2026. Referendums and the double majority.
143. NSF. 2026. NAIRR at two years.
144. European Commission. Checked 5 November 2026. AI factories.
145. UK Government. 2025; checked 5 November 2026. UK Compute Roadmap.
146. Epoch AI. Checked 5 November 2026. AI Chip Users: methodology.
147. Epoch AI / EPRI. 11 August 2025. Power demands of frontier AI training.