Running 100,000 GPUs at Once: The Texas Data Center Behind GPT-6 Astra

On September 3, 2026, OpenAI unveiled GPT-6 Astra, releasing it to the public the following day.

The benchmark scores immediately drew headlines: 97.6% on challenging math problem sets, 99.9% on ARC-AGI-3, and 100% on vulnerability discovery benchmarks. Naturally, speculation that “AGI has arrived” quickly followed.

This article is not about those claims. Instead, it looks at the underlying physical layer required to produce such a model. In short, what changed most this time around was not the algorithms, but the concrete and power delivery.

First: Is It Really AGI?

Let’s clarify up front: a high ARC-AGI score does not mean AGI has been achieved.

Despite having “AGI” in its name, ARC-AGI is simply the name of a benchmark test suite. It is an evaluation designed around novel visual and abstract reasoning patterns that humans solve easily but AI typically struggles with. Scoring 99.9% means the model “solved every problem on this test sheet,” not that it “thinks like a human.”

When a benchmark is saturated, it has reached the end of its useful life, and researchers build a new one. The name ARC-AGI-3 already signals that this is the third iteration.

That said, the performance leap is substantial. And the real driver behind that leap is clear.

① The Number: “100,000 GPUs”

A comment to reporters by Aidan Clark, VP of Research at OpenAI, captures the core of this release:

“This was overwhelmingly our largest training run to date. It marks the first time we’ve pretrained across more than 100,000 GPUs at the Texas Stargate site.“

Here, pretraining refers to the phase where the base model is trained from scratch on vast corpora of data. It is the longest and most capital-intensive stage of development.

Why is scaling to 100,000 GPUs so difficult? Why can’t you just rack 100,000 servers in a row?

Because you have to interconnect them so they act as a single machine.

Consider an analogy: having 100,000 people work independently in separate offices is straightforward. But what if all 100,000 must edit a single document simultaneously, reconciling edits in real time with everyone else? As headcounts grow, the communication overhead quickly consumes all the actual work time.

AI training runs into the exact same challenge. GPUs must continuously exchange intermediate computational states with one another. Consequently, the real bottleneck is not adding compute units, but how quickly those compute units can communicate.

We’ve explored related issues before — thousands of GPUs stalled waiting for storage I/O stems from the same family of scaling bottlenecks.

② Where Those 100,000 GPUs Live: Abilene

Abilene, Texas is a city of roughly 120,000 residents. It is home to the data center campus known as Stargate.

The facility is owned by Oracle and operated for OpenAI — a model where the cloud provider builds the infrastructure and shell, and the AI lab leases the compute capacity.

According to tracking data compiled by research group Epoch AI as of late July 2026, the numbers stand as follows:

H100-equivalent is a normalized metric. Because GPU architectures differ across generations, total compute is converted into the equivalent throughput of standard H100 units to allow fair comparison.

To put 421 MW into perspective: a typical nuclear reactor unit produces roughly 1,000 MW. This single data center campus draws nearly half the output of a full-scale nuclear plant.

③ The Construction Timeline

This timeline highlights the driving constraint behind modern frontier models:

May 2024      Site grading begins
Sep 2025      Building 1 live   (105 MW)
Dec 2025      Building 2 live   (211 MW)
May 2026      Buildings 3 & 4 live (421 MW)
Sep 2026      ← GPT-6 Astra training completed & launched
Nov 2026      Buildings 5–8 scheduled to go live

Buildings 3 and 4 were powered on in May; the model shipped in September.

In other words, the training run began the moment the facility went live. The release date was governed not by silicon availability or algorithmic breakthroughs, but by concrete curing times and switchgear installation schedules.

It took two full years to go from raw dirt to powering four buildings. Throughout that period, the software stack had to wait on civil and electrical engineering.

④ Hardware Architecture and Packaging

Epoch AI’s current accounting indicates the deployment consists of roughly 100,000 Nvidia B200s and 100,000 B300s — Nvidia’s Blackwell-generation AI accelerators.

An accelerator refers to specialized silicon optimized for tensor math. GPUs are the primary example and the term is often used interchangeably.

The rack design is notable. The public spec uses the GB200 NVL72 platform. Rather than a collection of independent servers, the entire rack acts as a single cohesive computing unit.

• Each rack houses 72 GPUs
• All 72 GPUs are linked via a high-bandwidth intra-rack copper interconnect, behaving as a single domain
• This topology significantly reduces the inter-node communication bottleneck at the rack level

The challenge is thermal dissipation. Packing 72 high-power GPUs into one rack is thermally equivalent to running dozens of industrial space heaters inside a standard closet. Air cooling is physically inadequate. Consequently, direct-to-chip liquid cooling is mandatory.

The ultimate heat rejection to the environment is handled by central chiller plants. The Abilene site currently runs 320 chillers providing 562.7 MW of cooling capacity. Once all 8 buildings are commissioned, this expands to 640 chillers and 1,125 MW.

The infrastructure stack scales as follows:

Silicon     B200 / B300 (Blackwell accelerators)
  ↓
Rack        GB200 NVL72 — 72 GPUs unified per rack
  ↓
Facility    Tens of thousands of GPUs per building across 8 buildings
  ↓
Power       421 MW (scaling to 843 MW)
  ↓
Cooling     320 chillers / 562.7 MW

Note: Target figures in marketing announcements (e.g., 450,000 GB200s, 1.2 GW) differ from Epoch AI’s observed operational figures (approx. 100,000 B200s and 100,000 B300s, 421 MW). Planned target capacity must be distinguished from currently energized capacity. Similarly, while PR materials mention “closed-loop liquid cooling,” operational filings list “air-cooled chillers” — reflecting the difference between the primary fluid loop at the chip level and the secondary heat rejection loop to ambient air.

⑤ Architectural Shifts on the Software Side

Hardware scaling alone does not tell the whole story. Astra also introduced structural changes, employing a looped transformer design (also referred to as recurrent depth).

Traditional transformers are built like multistory towers. Input tokens pass sequentially through Layer 1 up to the top layer, being refined at each step. Increasing reasoning depth historically meant stacking more physical layers, which increases parameter count, memory footprints, and serving costs.

The recurrent approach takes a different route: data is passed iteratively through the same set of layers multiple times. Looping through a 10-layer block 5 times yields the computational depth of 50 layers, but the parameter footprint remains that of a 10-layer model. Neither model size nor base parameter memory increases.

This design introduces a new trade-off: it reduces the interpretability of intermediate chains of thought. Because activations loop through shared weights, external inspection of latent reasoning steps becomes significantly harder, creating potential challenges for mechanistic interpretability and safety evaluations.

OpenAI Chief Scientist Jakub Pachocki stated that they cap the maximum recurrence iterations to maintain observability and bound compute latency.

⑥ At 100,000 GPUs, Hardware Failure Is Continuous

Finally, there is the operational reality of running at this scale — a topic familiar to infrastructure engineers.

When operating 100,000 enterprise GPUs continuously for months, component failures are not anomalies; they are continuous background events. Even with hardware boasting low annualized failure rates, a fleet of 100,000 units experiences multiple component faults every single day.

Because distributed training tightly synchronizes the entire cluster, a silent hang or crash on a single GPU can block or invalidate the entire training run. Unlike stateless web services, you cannot simply drop the unhealthy node from a load balancer pool and continue uninterrupted.

Silent data corruptions (SDCs) are even more insidious: hardware calculating mathematically incorrect values without raising an error interrupt. Left undetected, training continues, silently degrading weights and invalidating days of compute. These dynamics are fundamentally reshaping system troubleshooting methodologies.

To address this, frontier AI labs now maintain dedicated engineering roles focused solely on hardware diagnostics and automated failure isolation. Indeed, job postings across top AI labs frequently list site reliability and diagnostic positions, with several requiring full-time on-site presence in Abilene — the very location profiled here.

What Has Changed for Cloud Infrastructure?

Three key takeaways stand out:

First, the primary scaling bottleneck shifted from silicon allocation to facility readiness. The issue is no longer just securing GPU allocations, but securing energizable megawatt capacity and high-density shells to house them. The Abilene schedule is a textbook example.

Second, the role of cloud providers has evolved. Oracle’s delivery here was not ordinary IaaS virtualization, but engineering and leasing a utility-scale power and cooling facility wholesale. AI labs lease entire dedicated substations and mechanical infrastructure rather than spinning up generic clusters.

Third, physical facility engineering has returned to center stage. Substation interconnects, heat-exchange thermodynamics, rack power densities, and optical interconnects — domains abstracted away during the early cloud era — are once again primary architectural constraints, now operating at utility scale.

For cloud and systems engineers, the critical takeaway from GPT-6 Astra is not the headline benchmark percentages. It is recognizing that frontier model capability is directly bounded by switchgear, cooling loops, and concrete, and that expertise spanning both physical systems and distributed software is increasingly vital.

⚠️ Note on data sources: The metrics in this article draw from two distinct categories. Operational verified figures (compute capacity, power draw, chiller count, building footprint, and estimated capex) reflect tracking data published by Epoch AI as of July 28, 2026. Announced targets (e.g., 450,000 GB200s, 1.2 GW) stem from OpenAI and Oracle press releases and represent planned, future deployments rather than active state. Statements regarding pretraining scale were made by OpenAI VP of Research Aidan Clark to the press. Descriptions of looped transformer mechanics summarize public technical discussions; OpenAI has not released full architectural specifications. GPT-6 Astra was released on September 3, 2026, and independent evaluations remain ongoing.

Frequently Asked Questions

Does scaling to 100,000 GPUs yield 100,000× the performance?

No. Compute scaling does not yield linear performance gains due to distributed communication overhead. As node counts grow, the time spent synchronizing gradients and intermediate states across the fabric increases. Consequently, engineering competition focuses less on total chip volume and more on scaling cluster-wide bisection bandwidth to maintain high Model Flops Utilization (MFU). Architectures like the GB200 NVL72 exist specifically to solve this by creating ultra-high-speed domains at the rack level.

Why build in locations like Abilene, Texas?

Land availability, power prices, and transmission grid access dictate site selection. Facilities drawing hundreds of megawatts cannot be integrated into dense metro grids. Developers prioritize access to high-voltage transmission lines, fast utility interconnection queues, and low industrial electricity rates. While practical for infrastructure, high utility draw has made local grid impact and water consumption active community topics, prompting labs to hire dedicated local infrastructure and policy liaisons.

Why is liquid cooling necessary?

Thermal design power (TDP) for modern accelerator racks has exceeded the physical limits of forced-air convection. Packing 72 high-wattage GPUs into a single standard rack footprint generates heat densities that air cannot evacuate effectively, regardless of fan velocity. Liquid has a significantly higher volumetric heat capacity, enabling direct heat removal at the cold plate. However, it introduces operational complexity around fluid distribution units (CDUs), manifold plumbing, and leak mitigation.

Does this impact standard enterprise cloud strategies?

Most organizations will not operate clusters at this scale, but market availability is directly affected. Regional GPU shortages, reserved capacity constraints, and pricing trends across major cloud providers are heavily shaped by multi-hundred-megawatt allocations committed to frontier training clusters. When accelerator capacity is constrained in a given cloud region, the bottleneck is often local grid substation allocations rather than server chassis inventory.

Are looped transformers strictly superior to standard architectures?

They represent an engineering trade-off. The primary advantage is achieving deep recurrent computation without expanding parameter footprints or static memory requirements. The downside is reduced interpretability: internal reasoning steps are harder to isolate when intermediate representations pass cyclically through shared layers. OpenAI addressed this by capping recurrence bounds to balance reasoning depth with stability and observability.

How should claims of reaching AGI be evaluated?

Benchmark saturation should not be conflated with general intelligence. ARC-AGI is a standardized benchmark suite; achieving near-perfect scores indicates the test has been solved within its evaluation distribution. As with prior iterations, saturated benchmarks are retired in favor of harder suites (e.g., ARC-AGI-3). While recent models demonstrate significant capability gains, high scores on specific test suites do not equate to autonomous general intelligence, particularly while independent replication and evaluation remain in early stages.