xAI's Cortex 2, Meta's mega-clusters, and the $100B NVIDIA-OpenAI deal represent computing installations that dwarf anything on the official rankings. The supercomputing world hasn't reckoned with what that means.
The TOP500 list, published twice a year since 1993, has been the benchmark for who has the world’s most powerful computers. It’s a public ranking: you submit your Linpack score, the list gets sorted, national pride gets stoked or bruised accordingly. For three decades, this system worked because the most powerful computers were funded by governments and operated by national labs. Public money, public benchmarks, public rankings.
That era is ending. In 2026, the most powerful computing installations on Earth are private, proprietary, and completely absent from the TOP500. They could easily dominate the rankings, but their operators have no incentive to submit and every incentive not to.
xAI’s Cortex 2 supercluster is powering up its first section in the first half of 2026 at Tesla’s Giga Texas campus in Austin. Details are sparse, by design, but what’s known is that the cluster is built around NVIDIA H100 and H200 GPUs in quantities that put it in the running for the most powerful single computing installation in the world.
xAI’s first supercluster, Cortex (informally called “the Memphis cluster”), drew attention in 2024 for both its speed of deployment—going from empty building to operational supercomputer in 122 days—and its aggressive power consumption. Cortex 2 is bigger. Substantially bigger. Elon Musk described it as necessary for training Grok 3 and beyond, models that require compute budgets exceeding anything in the Grok 2 training run.
How powerful is Cortex 2? We don’t know precisely, because xAI isn’t telling anyone. But we can estimate. If the cluster contains 100,000+ H100 GPUs (a number consistent with xAI’s known procurement volumes), the theoretical peak FP64 performance exceeds anything on the current TOP500. If it’s been upgraded or supplemented with H200 or Blackwell hardware, the gap widens further.
NVIDIA and OpenAI announced a data center partnership valued at approximately $100 billion, with the first installations built around Vera Rubin-based systems. The scale of this deal is difficult to contextualize. One hundred billion dollars buys a lot of silicon. If even a fraction of that is deployed as a single coherent computing system, it would exceed Frontier’s performance by an order of magnitude or more.
The details of the partnership—how the compute is deployed, where the facilities are located, what the system architecture looks like—are commercial secrets. OpenAI isn’t submitting anything to the TOP500. Neither is NVIDIA. The benchmark that defined supercomputing for thirty years is irrelevant to the organizations building the world’s most powerful computers.
Meta’s AI infrastructure buildout is the least visible and potentially the largest. The company has committed $115–135 billion to AI infrastructure in 2026 alone, a number that dwarfs the total budget of every national lab in the United States combined over the same period.
Meta’s AI training clusters are distributed across multiple data centers, but the company has been consolidating toward larger, more tightly-coupled installations. Internal Meta publications describe training runs on clusters of 16,000+ GPUs running for weeks. The infrastructure being built in 2026 is designed for runs on clusters an order of magnitude larger.
None of this shows up on the TOP500. None of it is publicly benchmarked. None of it is available for external research. It’s computing power that exists entirely for Meta’s internal use, with the occasional publication of model weights (Llama) being the only public output.
The reasons are straightforward. Submitting to the TOP500 requires running the Linpack benchmark, which takes time on hardware that could be generating revenue. It also reveals system architecture details that companies consider competitive intelligence. And there’s no business benefit; nobody chooses a cloud provider based on TOP500 rankings.
But the absence of these systems from public benchmarking has consequences for the broader computing community.
Scientific benchmarking loses its reference point. When researchers say their algorithm achieves “X% of peak performance,” that peak is defined by benchmarked systems. If the actual peak is 10x higher and unmeasured, the field’s sense of what’s possible becomes distorted.
Public policy operates with incomplete information. Governments making decisions about HPC investment—how much to spend on national lab systems, what performance targets to set, how to regulate AI compute—are working with a TOP500 that dramatically understates the total installed base of supercomputing capacity. The TOP500’s aggregate performance might represent a third or less of the actual global supercomputing capacity when private clusters are included.
The talent pipeline shifts. Graduate students and early-career researchers can access national lab systems through allocation programs. They can’t access xAI’s Cortex 2 or Meta’s internal clusters. As the majority of the world’s compute moves behind corporate walls, the training ground for the next generation of computational scientists shrinks.
A rough accounting of private AI supercomputing capacity in 2026, based on public statements, procurement data, and facility power ratings:
xAI (Cortex + Cortex 2): 100,000+ GPUs, estimated aggregate performance exceeding 10 exaFLOPS FP16. Meta: multiple clusters totaling 300,000+ GPUs across several facilities, estimated aggregate exceeding 30 exaFLOPS FP16. Google: TPU v5 and v6 clusters across multiple data centers, combined capacity not publicly disclosed but estimated comparable to Meta. Microsoft/OpenAI: expanding Azure AI infrastructure plus dedicated OpenAI training clusters, with the $100B NVIDIA deal adding substantially. Amazon: Trainium-based clusters plus NVIDIA GPU deployments, capacity growing rapidly.
Compare this to the TOP500’s current aggregate: roughly 5–6 exaFLOPS of Linpack performance across all 500 systems. The private sector’s installed AI supercomputing capacity likely exceeds the entire TOP500 by a factor of 10 or more. And that gap is widening every quarter.
Beyond the Big Five hyperscalers, a second tier of companies is building substantial AI computing installations. Oracle is constructing what it calls a “flagship AI data center,” discussed on Bloomberg TV but with limited technical details. Nebius Group, which spun out of Yandex’s international AI business, is projecting 521% revenue growth in 2026 as an AI infrastructure provider. CoreWeave, Lambda, and a growing cohort of GPU cloud providers are building clusters in the tens of thousands of GPUs.
The aggregate capacity of this second tier is meaningful. Together, they may control more GPU compute than any single hyperscaler. And they’re the most likely to eventually submit to the TOP500 or equivalent rankings, since public benchmarks help them attract customers. The irony is that the companies most motivated to show off their computing power publicly are the ones with the least of it.
The TOP500’s relevance is fading not because of a flaw in its methodology but because the world it was designed to measure has fundamentally changed. When Jack Dongarra and his colleagues created the list in 1993, virtually all supercomputers were publicly funded and publicly operated. The list was comprehensive because the community it measured was open.
A few options:
A private-sector track. Create a confidential benchmarking program where companies can submit results that are aggregated (total AI supercomputing capacity) without revealing individual system details. This gives the community a realistic picture of global computing capacity without compromising competitive intelligence.
Workload-specific benchmarks. Linpack measures dense linear algebra performance. AI workloads don’t look like Linpack. MLPerf exists but has limited adoption at the hyperscaler level. A benchmark suite that captures the actual performance profile of large-scale AI training—including communication overhead, checkpoint time, and fault tolerance—would be more useful to the community and potentially more attractive to private-sector participants. The state of MLPerf benchmarking shows the gap between what is measured publicly and what is actually deployed at scale.
Power-normalized rankings. Given the energy constraints driving the industry, a benchmark that measures performance per watt at scale could attract participants who want to demonstrate efficiency leadership. The Green500 exists as a sub-ranking, but it’s treated as a sideshow. Making energy efficiency the primary ranking criterion would be timely.
None of these will happen quickly. The TOP500 has institutional momentum and genuine value as a longitudinal dataset spanning three decades. But pretending it still represents the world’s most powerful computers is increasingly fictional.
The age of the public supercomputer isn’t over. Frontier, Aurora, and El Capitan continue to do work that private clusters can’t or won’t do: open science that benefits everyone. But the center of gravity in raw computing power has shifted decisively to the private sector. The supercomputing community needs to acknowledge that shift and figure out what public supercomputing’s role is in a world where the biggest machines are behind locked doors.
