Note: This is a research note supplementing the book Unscarcity, now available for purchase. These notes expand on concepts from the main text. Start here or get the book.

Compute Clusters: The Factories of the Intelligence Age

How AI training facilities became the oil refineries of the 21st century - and why your grandchildren might receive a compute allocation alongside their birth certificate.

What Is a Compute Cluster, Really?

“Compute cluster” sounds impressively technical but obscures something simple. A compute cluster is just a lot of computers working together on the same problem. The magic is in the “working together” part.

When you train a large language model like GPT-4 or Llama 4, the task is far too massive for any single computer, even an obscenely powerful one. You need to split the work across thousands of processors that constantly communicate, share results, and coordinate their efforts. If one machine calculates something, every other machine needs to know about it immediately. The lag between them must be measured in microseconds, not seconds.

Building a pyramid is a useful analogy. One person with a wheelbarrow would take millennia. A thousand people need to coordinate: who’s carrying what, where’s the next block going, don’t drop that on Steve. A compute cluster is that coordination problem, solved at the speed of light.

The processors doing this work are Graphics Processing Units (GPUs), originally designed to render video game graphics but repurposed for AI because they excel at doing many simple calculations simultaneously. If a traditional CPU is a brilliant surgeon performing delicate operations one at a time, a GPU is a factory floor with thousands of workers each doing one small task very quickly.

The Current State of Play: Welcome to the GPU Arms Race

In 2024-2025, we crossed into new territory. The compute clusters training frontier AI models became genuinely staggering in scale:

xAI’s Colossus: From 100,000 to 500,000+ GPUs

Elon Musk’s AI company, xAI, built what was briefly the world’s largest AI supercomputer in a converted Electrolux factory in Memphis, Tennessee. The timeline was absurd: from decision to operational in 122 days, when similar projects typically take years. By September 2024, it was running 100,000 Nvidia H100 GPUs.

Then they doubled it, in 92 days, and kept going.

By the end of 2025, Colossus had reached 200,000 GPUs: a mix of 150,000 H100s, 50,000 H200s (141GB of HBM3e memory each, nearly double the H100’s 80GB), and Blackwell-generation chips. In January 2026, xAI purchased a third building, expanding the site to 2 gigawatts of total power capacity and 555,000 Nvidia GPUs purchased for approximately $18 billion. As of early 2026, the combined Memphis sites operate 400,000-550,000 GPUs (primarily Blackwell GB200/GB300 series in the newer buildings), drawing 300 MW from the grid with a target of 1 GW. The ultimate goal remains one million GPUs.

The buildout kept accelerating into spring 2026. Colossus 2 alone runs 200,000 GPUs at 140 MW, reported as the fastest-growing AI data center in the world. Colossus 1 crossed 230,000 GPUs, including more than 30,000 of Nvidia’s newest GB200 Blackwell chips. xAI then filed a $659 million permit to expand Colossus 2 with a new four-story, 312,000-square-foot building, and grabbed a third site across the state line in Southaven, Mississippi, nicknamed “MACROHARDRR.” Hit the one-million-GPU target by late 2026 and the complex pushes past 3 gigawatts.

Then came the twist: the fortress became a rental property. xAI was folded into SpaceX, which went public on June 12, 2026 at $1.77 trillion, the largest IPO on record, and its S-1 revealed that the biggest users of Colossus are xAI’s competitors. Anthropic leases the entire output of Colossus 1 (about 300 MW) for $1.25 billion a month through May 2029. Google pays $920 million a month from October 2026 through June 2029 for roughly 110,000 GPUs, which a Google Cloud spokesperson described as “bridge capacity” because Gemini demand outran its own buildout. That’s about $26 billion a year in rent, from two of the best-funded AI companies on Earth, to a direct rival. The builders are becoming landlords, and the tenants are giants.

This is no R&D project. It is industrialized intelligence production at a scale that didn’t exist two years ago. It also draws on a shared commons: when xAI’s power demands materialized, Memphis residents raised concerns about grid reliability and electricity prices, and the Tennessee Valley Authority had to scramble to keep Colossus fed without browning out everyone else. The electrons flowing into a private training cluster are electrons not flowing into homes, hospitals, and schools.

Meta’s Infrastructure: Millions of GPUs

Meta started the AI boom by accumulating 350,000 H100 GPUs by end of 2024. That was just the beginning. In early 2026, Meta and Nvidia announced a multiyear partnership covering “millions” of Nvidia’s Blackwell and upcoming Rubin GPUs, one of the largest single infrastructure commitments in semiconductor history. A week later, Meta struck a separate deal with AMD for up to 6 gigawatts’ worth of GPUs. The company has since raised its 2026 capital expenditure guidance to $130 to $145 billion, roughly double the $72.2 billion it spent in 2025. Meta is also building Hyperion, a massive AI campus in Louisiana with 7.5 gigawatts of gas-fired power capacity. And on July 1, 2026, Bloomberg reported that Meta is building a cloud business to sell its excess AI capacity, reportedly called Meta Compute, offering both raw GPU cycles and hosted model access in direct competition with AWS, Azure, and Google Cloud. Wall Street’s verdict: the stock jumped about 10% in a day.

Four weeks later the market priced the same strategy in reverse. Meta’s second-quarter results, reported July 29, 2026, showed $31.1 billion spent on data centers, servers, and chips in a single quarter, up 83% year over year, against $31.9 billion of operating cash flow. Free cash flow came in at $784 million, down from $8.55 billion in the same quarter a year earlier, and the stock fell as much as 10% after hours. On the call, Zuckerberg said Meta is “getting a lot of offers for compute at a significant premium over what we paid for it” and is turning them down for now, on the theory that selling intelligence built on the compute beats renting the compute out. Sit with what that premium implies. Buyers will pay above cost for someone else’s already-energized capacity, because a chip order clears in months while a gigawatt-scale grid interconnection takes years. Money can buy silicon. It cannot buy a place in the queue.

Their Llama 3 model trained on a cluster of 16,384 H100s for 54 days, during which they experienced 148 interruptions from faulty GPUs and 72 from memory failures. This is rocket science that involves replacing engines mid-flight.

Oracle’s Zettascale Vision

Oracle is now taking orders for what they call the first “zettascale” AI supercomputer: up to 131,072 GPUs in a single cluster. For context, that is more than three times the computing capacity of Frontier, the world’s fastest traditional supercomputer.

The Stargate Project: $500 Billion Meets Reality

Announced in January 2025 with a White House press conference, the Stargate Project represents perhaps the most ambitious AI infrastructure buildout ever conceived. SoftBank, OpenAI, Oracle, and MGX pledged up to $500 billion over four years, with $100 billion deployed immediately.

By early 2026, reality has set in. The flagship site in Abilene, Texas is operational but facing challenges: winter weather knocked buildings offline, and OpenAI has decided against further expansion at Abilene, instead spreading capacity across more than half a dozen sites nationwide. OpenAI still doesn’t own any data centers, relying on Oracle, Microsoft, and Amazon for capacity. Microsoft has taken over two additional Abilene facilities, bringing the complex to ten buildings with roughly 2.1 gigawatts of projected capacity. The first gigawatt of Nvidia GPU systems is targeted for the second half of 2026, though experts warn the timeline is tight. The map keeps growing regardless: in June 2026, OpenAI and Oracle added a new Stargate campus in Michigan.

What’s Inside These Clusters: The Hardware

The Workhorse: Nvidia H100

The Nvidia H100 (codenamed “Hopper”) was the chip that defined the 2023-2024 AI boom and remains widely deployed. Key specifications:

80GB of HBM3 memory (the fastest memory type available at launch)
3.35 terabytes per second memory bandwidth (how fast data moves)
4 petaflops of AI compute (a petaflop is a quadrillion calculations per second)
~$25,000-40,000 per chip new, with used units trading at $15,000-25,000

The H100 was genuinely scarce in 2023-2024, with AI companies hoarding them like gold bars. By 2026, cloud rental prices have dropped 64-75% from their peak, though a recent surge in inference demand (driven partly by AI coding tools) has pushed H100 rental rates back up roughly 40% from their October 2025 lows.

Blackwell: The Current Generation

Nvidia’s Blackwell architecture shipped at scale through 2025 and drove Nvidia to a record $215.9 billion in fiscal year 2026 revenue. The B200 GPU represents a generational leap over Hopper:

192GB of HBM3e memory (2.4x the H100)
8 terabytes per second bandwidth (2.4x the H100)
20 petaflops of AI compute (5x the H100)
~$30,000-40,000 per chip

The GB200 NVL72 system connects 72 Blackwell GPUs to act as a single massive processor with 1.4 exaflops of AI performance. That’s 1.4 quintillion calculations per second. For perspective, the human brain is estimated to perform about one exaflop, so this rack of GPUs matches 1.4 human brains in raw computational throughput.

The Next Wave: Vera Rubin

In January 2026, Nvidia’s Vera Rubin platform, the successor to Blackwell, entered commercial production. The Vera Rubin NVL72 rack links 72 Rubin GPUs with 36 Vera CPUs using NVLink 6 switches. The numbers are striking: 3.5x faster than Blackwell for training, 5x faster for inference, up to 50 petaflops of performance, and 10x the inference throughput per watt at one-tenth the cost per token. At GTC 2026, Jensen Huang projected at least $1 trillion in orders for Blackwell and Vera Rubin chips through 2027, double the $500 billion he cited just six months earlier. The next architecture after Rubin, codenamed Feynman, is already in development targeting TSMC’s A16 process node.

The Networking: NVLink and InfiniBand

Raw GPU power means nothing if the chips can’t talk to each other fast enough.

NVLink is Nvidia’s proprietary interconnect for GPU-to-GPU communication within a single server or rack:

1.8 terabytes per second of bidirectional bandwidth (the latest version)
7x faster than PCIe Gen 5 (the standard computer connection)
Enables GPUs to share memory directly, as if they were one chip

InfiniBand connects servers across the data center:

400-800 Gb/s per port (with Quantum-X800 InfiniBand now shipping)
Under 100 nanoseconds latency (a nanosecond is a billionth of a second)
RDMA capability (Remote Direct Memory Access - GPUs can read each other’s memory without involving the CPU)

The typical architecture: NVLink connects GPUs within a node, InfiniBand connects nodes across the cluster. Together, they make thousands of GPUs behave like one giant processor.

The Power Problem: These Things Are Hungry

The compute cluster buildout collides with physical reality: these facilities consume obscene amounts of electricity.

According to the International Energy Agency:

Global data center electricity consumption hit 415 TWh in 2024 (~1.5% of global electricity), rising to an estimated 550 TWh in 2026 (~2%)
U.S. data centers alone consumed 183 TWh in 2024, about 4% of American electricity, and on course for 7-12% by 2028
By 2030, global data center consumption is projected to reach 945 TWh (~3% of global electricity), equivalent to Japan’s entire electricity demand

AI specifically consumes 10-20% of current data center energy, but that fraction is rising rapidly, potentially to 35-50% by 2030. Data center electricity demand is growing at roughly 15% per year, more than four times faster than all other sectors combined.

The xAI Colossus site in Memphis now draws 300 MW from the grid and is targeting 1 GW. Meta’s Hyperion campus in Louisiana has contracted 7.5 GW of gas-fired power, a 30% increase to the state’s entire grid capacity.

Google’s environmental report, published July 2, 2026, shows what this looks like on one company’s meter: electricity use up 37% in 2025, the largest single-year increase in its history, with data centers alone consuming 42 million MWh. That is one firm drawing roughly a mid-size country’s annual power. Google’s own explanation for why it expects the curve to steepen is the sentence planners should be reading twice: the buildout is accelerating faster than the grid is decarbonizing.

This is more than an engineering problem. It is a civilizational constraint.

The Fusion Connection

The tech industry’s obsession with fusion energy is grounded in arithmetic, not PR. When you’re planning to run data centers that individually consume gigawatts of power, you need energy sources that scale.

Microsoft has signed a deal with Helion to purchase power from a fusion reactor by 2028. Helion broke ground on its Orion commercial plant in 2025 and has raised over $1 billion in total funding. OpenAI has since signed its own deal with Helion for 5 GW by 2030, scaling to 50 GW by 2035. Google signed a 200 MW fusion PPA with Commonwealth Fusion Systems, which raised $863 million in 2025 and expects its SPARC prototype to operate in 2026. Total private fusion investment has reached roughly $10 billion.

The timeline is aggressive but not arbitrary. Current grid infrastructure simply cannot support the projected AI data center buildout. Something has to give. Either AI development slows, or we find new energy sources.

This is the fuel component of the Unscarcity tripod (The Brain, The Body, The Fuel). Fusion and AI are symbiotic necessities, not merely adjacent technologies.

The Hyperscaler Investment: Hundreds of Billions

The capital flowing into AI infrastructure defies historical precedent:

Company	2025 Actual / 2026 Guidance	Focus
AWS	$132B (2025) / ~$200B (2026)	Cloud AI infrastructure
Google	~$75B (2025) / $175-185B (2026)	Cloud, Gemini training
Meta	$72.2B (2025) / $130-145B (2026)	Llama training, social AI
Microsoft	~$80B (2025) / ~$150B pace (2026)	Azure, OpenAI/Stargate
Oracle	Stargate partner	Major infrastructure expansion

Just four companies (Amazon, Microsoft, Alphabet, and Meta) are projected to spend about $630 billion on data centers and AI chips in 2026 alone, according to Morgan Stanley. Include the top 11 cloud and infrastructure providers (Oracle, CoreWeave, etc.) and total CapEx hits $811 billion. By 2030, cumulative AI data center capital expenditure is projected to exceed $5.2 trillion.

These numbers are difficult to contextualize. The entire Apollo program cost about $280 billion in today’s dollars. The 2026 AI infrastructure spend alone is more than double that. The $630 billion from just four companies amounts to roughly 2.2% of U.S. GDP.

The Geopolitics: Compute Is the New Oil

Sam Altman has called compute “the currency of the future”, “possibly the most precious commodity in the world.” Jonathan Ross of Groq echoes the sentiment: “Compute is the new oil.”

The phrase is not hyperbole. The concentration is staggering:

The United States controls ~75% of the world’s AI supercomputing capacity
China holds ~15% (and falling, due to tightened export restrictions)
Nvidia commands ~85% of the GPU market overall, with 80%+ of AI training chips (down slightly from 92% in early 2025 as AMD reaches ~7%)
TSMC manufactures ~90% of the world’s advanced chips, with $52-56 billion in 2026 CapEx and up to 10 new fabs under construction

The U.S.-China semiconductor conflict is, in effect, a war over the means of intelligence production. The CHIPS and Science Act allocated over $52 billion to incentivize domestic chip manufacturing. TSMC is now planning up to 12 fabs and four advanced-packaging facilities in Arizona, far beyond the original two-fab plan. Equipment installation for its N3 fab begins mid-2026, with 2nm/A16 chips to follow. The explicit goal is to reduce dependence on Taiwan, which sits 100 miles from mainland China.

Export controls have made China a marginal producer of AI chips. As of April 2026, the U.S. has tightened controls further, and Nvidia’s market share in China has dropped from 95% to roughly 55%, with Huawei gaining ground. DeepSeek founder Liang Wenfeng stated bluntly: “Money has never been the problem for us; bans on shipments of advanced chips are the problem.”

Yet China is responding. SMIC continues advancing domestic chip production. Beijing’s “Big Fund” pours billions into semiconductor development. A technological decoupling is underway, creating two increasingly incompatible AI ecosystems, American and Chinese, competing for global influence.

The Taiwan Question

The uncomfortable truth: if China blockades or invades Taiwan, the global tech economy collapses overnight. Modern militaries would face an immediate semiconductor drought. Every iPhone, every AI model, every advanced weapon system depends on chips that mostly come from one small island.

This is why TSMC is expanding globally: to Japan (with Japanese government subsidies, already upgrading its Kumamoto fab from 6nm to 3nm), to Arizona (potentially 12 fabs worth over $100 billion), to Germany. Geographic diversification has become a national security imperative, well beyond ordinary business strategy.

Terafab: Owning the Whole Stack

The vertical-integration logic reaches its extreme with Terafab. On March 23, 2026, Engadget reported that Elon Musk announced a joint venture between Tesla, SpaceX, and xAI to build what he called the largest chip manufacturing facility ever constructed, in Austin, Texas, targeting one terawatt of annual compute production. The stated purpose is to produce custom silicon for Tesla vehicles, SpaceX spacecraft, xAI’s training clusters, and Optimus humanoid robots, and to stop buying from TSMC and Samsung entirely.

The announcement omitted a production timeline, a confirmed chip architecture, and an investment figure, which is most of the business case. Building a leading-edge fab is among the hardest industrial undertakings humans attempt: TSMC spent more than $30 billion on its Arizona fabs and they still aren’t fully operational, and the entire global semiconductor industry currently produces chips consuming roughly 300-400 gigawatts at peak load. A single facility claiming a terawatt of yearly output would, in theory, produce more compute capacity per year than the world currently uses. The energy demand alone is daunting. TSMC’s Taiwan facilities use about 5% of that country’s total power, and a single EUV lithography machine draws roughly one megawatt.

Whether or not Terafab is built on Musk’s stated timeline (his track record runs late but directionally correct), the structure is the point. One ecosystem would design the chips, manufacture them, install them in cars and robots and rockets, train AI on them, and deploy the result into the physical world, with no congressional oversight of the architectures, no independent audit of the models baked into the silicon, and no public review of the supply chain. Compute is becoming the new oil, and Terafab is an attempt to own the refinery, the pipeline, and the gas stations at once. The deeper concentration problem is taken up in Infrastructure Libertarianism and The Iron Law of Oligarchy.

Why Individual Chips Don’t Matter (And Why Clusters Do)

A single H100 is impressive but useless for frontier AI training. The models are simply too large. GPT-4 is rumored to have around 1.7 trillion parameters. Llama 3’s largest version has 405 billion. Each parameter needs to be stored, updated, and communicated.

The math works like this:

A 1 trillion parameter model needs ~2 terabytes just to store the weights in half-precision floating point
During training, you also need to store gradients and optimizer states, roughly 16 bytes per parameter
That’s 16 terabytes of memory for a 1 trillion parameter model
A single H100 has 80GB of memory

You need at minimum 200 H100s just to hold the model, plus additional GPUs for batch processing. In practice, training requires thousands to tens of thousands of GPUs running for weeks or months.

This is why clusters matter. The core technological achievement is the ability to coordinate tens of thousands of processors on a single training run, with minimal communication latency, maximal bandwidth, and reliable fault tolerance. Any individual chip is impressive on its own, but the orchestra is what transforms the field.

The Coming Democratization (Maybe)

The Unscarcity thesis becomes concrete at this point.

Right now, frontier AI training is confined to about a dozen organizations worldwide: OpenAI, Anthropic, Google DeepMind, Meta, xAI, Mistral, a handful of Chinese labs, and some well-funded startups. The capital requirements (billions of dollars in GPUs, power, and engineering talent) create natural barriers.

But three forces could democratize access:

1. Inference Is Cheaper Than Training

Training GPT-4 required months on a massive cluster. Running inference (getting answers from the trained model) is far cheaper. Cloud providers now offer API access to frontier models for pennies per query. The capability is increasingly accessible even as the means of production remains concentrated.

2. Smaller Models Are Getting Better

Distillation techniques, efficient architectures, and better training data mean smaller models can approach the performance of larger ones. Llama 3’s 8 billion parameter model outperforms GPT-3.5 (175 billion parameters) on many benchmarks. The floor is rising.

3. Decentralized Compute Networks

Emerging networks like Aethir aggregate idle GPU capacity globally. The idea: if a million people each contribute a small amount of compute, you get a distributed supercomputer. The technology is nascent, and the concept is sound.

Connection to the Unscarcity Vision: Universal Basic Compute

This brings us to Universal Basic Compute (UBC): the idea that every citizen should receive a guaranteed allocation of AI processing capacity.

If compute is becoming the means of production, the modern equivalent of land in an agrarian economy, then distributing compute is distributing economic agency. A citizen with a compute allocation can:

Run personal AI agents that manage life administration
Participate in pooled cooperatives for larger projects
Delegate their allocation to Mission Guilds in exchange for services
Create, research, or build using the same tools as major corporations

The Foundation layer in the Unscarcity framework could eventually provision compute the way it provisions housing or food: as infrastructure for dignified existence, a baseline rather than a luxury.

None of this is fantasy. Research institutions already allocate compute quotas. National cloud initiatives in Saudi Arabia, UAE, and Singapore are building sovereign AI infrastructure. The concept of “compute as public utility” is emerging in policy discussions.

But achieving UBC requires:

Sufficient total compute - currently a constraint, but Blackwell, Vera Rubin, and future generations are multiplying capacity at 10x per generation
Distribution infrastructure - cloud platforms that can allocate and meter compute fairly
Accessible interfaces - so “using compute” becomes as intuitive as “using electricity”
Governance frameworks - to prevent concentration, corruption, and hoarding

The Diversity Guard mechanism addresses governance. The Civic Service program includes AI orchestration training. The EXIT Protocol negotiates with existing power holders for gradual transition.

What to Watch For

2026-2027

Vera Rubin GPUs ramp to full production (H2 2026), followed by Feynman architecture (2028)
Stargate deploys first gigawatt of capacity across multiple U.S. sites
Hyperscaler CapEx surpasses $800 billion annually
The compute landlord market matures: Meta Compute, SpaceX-xAI leases, and CoreWeave compete to rent capacity - watch whether pricing trends toward commodity utility or gatekept rent
Continued export restriction tightening; two-ecosystem decoupling accelerates
TSMC begins N3 production in Arizona, with 2nm to follow
States and Congress move to reallocate the grid costs of data center demand onto operators (who pays for AI’s electricity)

2027-2030

First commercial fusion reactors potentially come online (Helion/Microsoft 2028, CFS/Google early 2030s)
AI-specific chip architectures (custom silicon from Meta, Amazon, Google) mature alongside Nvidia
Inference compute demand surpasses training compute
UBC pilot programs in selected jurisdictions

2030+

Fusion energy begins meaningfully powering data centers
Foundation-layer compute infrastructure in early deployments
Cumulative AI infrastructure spend exceeds $5 trillion
The contours of “who controls intelligence production” become clear

The Stakes

Something specific is being decided in the GPU warehouses of Memphis, Abilene, and undisclosed locations worldwide:

Who will control the means of intelligence production?

If the answer is “a handful of corporations and nation-states,” we get the Star Wars trajectory: technological feudalism with better graphics. The abundance AI creates concentrates at the top. Everyone else becomes economically irrelevant but biologically alive.

If the answer is “distributed infrastructure with universal access,” we get something closer to the Unscarcity vision, where compute is a public utility, AI capability is broadly held, and the Foundation layer provides dignified existence while the Frontier layer rewards contribution.

The race to build compute clusters is a commercial competition, but it is also the construction of the infrastructure that will determine how intelligence is distributed across human civilization.

The factories of the Intelligence Age are being built right now. The question is whether they’ll produce liberation or lock-in.

AI Compute Clusters 2026: The $800B GPU Arms Race