Anthropic’s Colossus2 Bet: $15 Billion a Year, GB200 Chips, and the Race to Own AI’s Physical Infrastructure

The numbers emerging from Anthropic this week would make even hardened Silicon Valley CFOs blink twice. On Wednesday, the company told investors it is on track to more than double its quarterly revenue to approximately $10.9 billion [2]. That is a staggering figure for a firm that, until recently, burned cash at a rate that made venture capitalists nervous. But the real story—the one explaining how Anthropic plans to sustain that trajectory—lies buried in a different set of documents. According to SpaceX’s long-awaited IPO filing, Anthropic is paying an eye-watering $15 billion per year to access data center capacity from Elon Musk’s rocket company [3]. Now, a new piece of the puzzle has clicked into place: Anthropic is expanding into a massive new cluster called Colossus2, powered by NVIDIA’s next-generation GB200 superchips [1].

This is not just another infrastructure expansion. It is a declaration of war on the entire prevailing model of AI compute. By moving to Colossus2 and committing to the GB200 architecture, Anthropic signals that it believes the future of frontier model training belongs to monolithic, ultra-dense superclusters—and that the company will spend like a nation-state to secure its place at the top. The implications ripple outward through supply chains, competitor strategies, and the increasingly tangled relationship between AI labs and the aerospace-industrial complex.

The Architecture Behind The Model: What Colossus2 Actually Is

Let’s start with the hardware, because the technical details are anything but trivial. The original Colossus cluster, built by xAI in Memphis, was already a marvel of rapid deployment and raw scale. Colossus2, as the name implies, is a second-generation facility—and the decision to outfit it with GB200 chips represents a bet on a specific architectural philosophy [1].

The GB200 is not merely a faster GPU. It embodies NVIDIA’s push toward what the company calls “superchips”: a tightly integrated package pairing Grace ARM-based CPUs with Blackwell-architecture GPUs, connected via a high-bandwidth, low-latency NVLink-C2C interconnect. For Anthropic, this matters because training models like Claude—which already rival and in some benchmarks surpass GPT-4 and Gemini—requires not just raw FLOPs but coherent memory access across thousands of accelerators. The GB200’s unified memory architecture allows the CPU and GPU to share a single address space, eliminating the traditional PCIe bottleneck that has historically plagued distributed training at scale.

The sources do not specify the exact size of Colossus2. We do not know whether it will house 10,000 or 100,000 GB200 units. But the $15 billion annual compute bill revealed in SpaceX’s IPO filing gives us a powerful clue [3]. If Anthropic pays that much for access to data centers—and SpaceX is just one of its providers—then Colossus2 must operate at a scale dwarfing anything previously built by a single AI company outside of Microsoft’s Azure supercomputers. The GB200 is not cheap; NVIDIA’s pricing for the superchip, while not publicly listed, is widely estimated at $30,000 to $50,000 per unit in volume. Multiply that by tens of thousands, add networking, cooling, power, and real estate, and $15 billion starts to look like a floor, not a ceiling.

This is where the editorial board’s announcement becomes genuinely disruptive [1]. By committing to Colossus2 now, Anthropic effectively locks in a multi-year supply of the most advanced silicon on the planet. Given that NVIDIA’s GB200 ramp is already oversubscribed by hyperscalers like Google, Amazon, and Microsoft, any allocation to a third-party AI lab represents a significant strategic concession from the chipmaker. It suggests that Anthropic has either paid a massive premium or negotiated exclusivity terms giving it priority access during critical early production runs.

The Financial Stakes: From Burning Cash to Printing It

The revenue numbers from Anthropic are almost as astonishing as the compute costs. The company told investors it will more than double revenue to around $10.9 billion in its second quarter [2]. To put that in perspective: Anthropic was founded in 2021. OpenAI took roughly eight years to reach a $10 billion annualized revenue run rate. Anthropic appears to be doing it in five, and accelerating.

But here is the tension every analyst is grappling with. If Anthropic pays $15 billion per year just for data center access from one provider [3], and its entire quarterly revenue is only $10.9 billion [2], the math does not immediately work. The company still spends more on compute than it earns—unless the revenue figure grows so fast that the second half of 2026 flips the equation entirely. The TechCrunch report explicitly says Anthropic is “about to have its first profitable quarter” [2], implying that the cost structure is finally coming into alignment. That could mean one of three things: (1) revenue growth is outpacing compute spend, (2) the $15 billion figure is amortized across multiple years and includes non-cash considerations, or (3) Anthropic has negotiated deeply discounted rates from SpaceX in exchange for equity or strategic partnership.

The Wired report, drawing from the SpaceX IPO filing, is careful not to specify the exact terms [3]. But the mere existence of a $15 billion annual commitment changes the narrative around Anthropic’s financial health. Historically, the company has been viewed as the safety-conscious, slower-moving alternative to OpenAI—the lab that prioritized alignment over speed. That reputation now collides with a spending profile suggesting Anthropic is willing to bet the entire company on scaling as fast as physically possible.

The SpaceX Connection: When AI Labs Become Aerospace Tenants

The most surreal element of this story is the landlord. SpaceX—a company whose primary business involves launching rockets, building satellites, and, as of its IPO filing, preparing to become a publicly traded entity—is now one of the largest providers of GPU compute on Earth [3]. The Wired report reveals that the long-awaited SpaceX IPO documents included details about a “lucrative deal to lend GPUs to a major AI rival” [3]. That rival is Anthropic.

This is not a simple real estate transaction. SpaceX has been building out its Starlink data center infrastructure for years, primarily to handle the massive telemetry and networking demands of its satellite constellation. But data centers designed for low-latency satellite routing are not necessarily optimized for AI training workloads, which require extreme power density, liquid cooling, and high-bandwidth interconnects. The fact that SpaceX has repurposed or expanded its facilities to accommodate Anthropic’s GB200 clusters suggests that the compute shortage is so acute that AI labs are turning to any organization with available power and cooling capacity—even if that organization’s core competency is building spaceships.

The strategic implications are profound. By colocating with SpaceX, Anthropic gains access to infrastructure that is geographically distributed, physically secure, and potentially powered by renewable energy sources that SpaceX uses for its launch and manufacturing operations. But it also creates a dependency that regulators may eventually scrutinize. If SpaceX becomes the dominant compute provider for frontier AI, and if SpaceX is also launching the satellites that will eventually provide global broadband, the concentration of power in a single corporate ecosystem—one already controlled by Elon Musk—raises questions that the Verge’s coverage of the Anthropic-OpenAI political rivalry only begins to touch on [4].

The Political Dimension: AI Labs at War, Super PACs, and the Midterms

Speaking of politics: the Verge’s Regulator newsletter this week highlighted that Anthropic and OpenAI are taking their long-simmering feud into the midterm elections [4]. The piece, written by Alex Bores, notes that both companies are deploying super PACs and lobbying resources to influence the 2026 congressional races. This is not a sidebar to the Colossus2 story—it is a direct consequence of it.

When you spend $15 billion a year on compute [3] and project $10.9 billion in quarterly revenue [2], you are no longer a startup. You are a systemic institution. And systemic institutions attract regulatory attention. Anthropic has historically positioned itself as the responsible actor in the AI safety debate, contrasting its “constitutional AI” approach with OpenAI’s more aggressive deployment strategy. But the Colossus2 expansion, combined with the SpaceX deal, suggests that Anthropic is playing the same scale game as everyone else—just with better PR.

The midterm spending [4] likely aims to shape the regulatory framework governing AI infrastructure. If the U.S. government decides to impose export controls on advanced chips, or to mandate that certain training runs must occur on domestic soil, the companies that have already locked in domestic compute capacity—like Anthropic, with its Colossus2 and SpaceX arrangements—will have a massive advantage. The super PAC money is an insurance policy, ensuring that the rules of the game favor incumbents with existing infrastructure.

What the Mainstream Media Is Missing: The GB200 Supply Chain Squeeze

Here is the hidden risk that most coverage glosses over. The GB200 is not just a chip; it is a system. It requires specialized cooling, custom server designs, and networking gear still in early production stages. NVIDIA has ramped up its supply chain for months, but demand from hyperscalers alone is enough to consume the entire output of TSMC’s CoWoS-L packaging capacity through 2027. If Anthropic is getting GB200 units for Colossus2, someone else is not.

The likely losers are smaller AI labs and enterprise customers hoping to get their hands on Blackwell-generation hardware for fine-tuning or inference. NVIDIA has historically prioritized its largest customers—the hyperscalers and frontier labs—during new product launches. By securing GB200 allocation for Colossus2, Anthropic effectively crowds out the mid-market. This will accelerate the consolidation of the AI industry into a handful of vertically integrated players who control their own compute, while everyone else rents capacity at increasingly exorbitant rates.

There is also a geopolitical angle. The GB200 is subject to U.S. export controls, and any cluster of this size would be a prime target for espionage or sabotage. Colossus2’s physical security, network architecture, and supply chain resilience will need to be military-grade. The fact that Anthropic is building this with SpaceX—a company that already operates under strict ITAR (International Traffic in Arms Regulations) compliance—suggests that the cluster may be treated as a national security asset from day one.

The Bottom Line: A Bet That Changes the Industry

Anthropic’s expansion to Colossus2 with GB200 chips is not a routine infrastructure upgrade. It signals that the company believes the era of renting compute from cloud providers is ending, and that winners in AI will own their hardware outright. The $15 billion annual payment to SpaceX [3] and the projected $10.9 billion quarterly revenue [2] paint a picture of a company spending aggressively to build a moat that no competitor can easily cross.

But the risks are equally enormous. If the GB200 ramp faces delays—and NVIDIA has a history of late-stage silicon issues—Colossus2 could come online months behind schedule, burning cash without producing models. If the midterm elections shift the regulatory landscape to restrict data center construction or energy consumption, Anthropic’s fixed-cost commitments could become liabilities. And if OpenAI or Google respond with even larger clusters—which they almost certainly will—the arms race could spiral into a level of capital expenditure that even $15 billion a year cannot sustain.

For now, though, Anthropic has the momentum. It has the revenue growth, the hardware allocation, and the political infrastructure to defend its position. Colossus2 is the physical manifestation of a company that has decided to stop being the thoughtful alternative and start being the dominant force. Whether that transformation preserves the safety-first ethos that defined Anthropic’s early years—or whether it gets crushed under the weight of its own ambition—is the question that will define the next chapter of the AI industry.

---

References

[1] Editorial_board — Original article — https://xcancel.com/nottombrown/status/2057194829986300375

[2] TechCrunch — Anthropic says it’s about to have its first profitable quarter — https://techcrunch.com/2026/05/20/anthropic-says-its-about-to-have-its-first-profitable-quarter/

[3] Wired — SpaceX IPO Filing Reveals Anthropic Is Paying $15 Billion a Year to Access Its Data Centers — https://www.wired.com/story/spacex-ipo-anthropic-compute-finances-risks/

[4] The Verge — Anthropic and OpenAI take their beef to the midterm elections — https://www.theverge.com/column/934684/anthropic-openai-super-pac-beef-alex-bores