The negative headlines around OpenAI, including a contested revenue miss and a user count tracking at 900 million against a stated target of 1 billion, are obscuring the more important variable for enterprise positioning. OpenAI continues to be a leader in raw compute capacity at exactly the moment Anthropic is bottlenecked on compute, and capacity is what determines whether a model can win and keep enterprise workflows. Once a large enterprise builds an automated workflow around a specific model, the switching cost is real, because tasks are chunked and chained and a single missed window from a compute shortage breaks the downstream sequence. If Claude cannot serve a request because it is out of compute, the enterprise routes that workflow to OpenAI or xAI and the revenue effectively leaves Anthropic in perpetuity. Large enterprises will increasingly contract directly with AWS or Azure for guaranteed compute and then choose the model on top, which separates the platform decision from the chip decision. Middle market buyers will not have that privilege and will be price-takers on whatever compute their AI platform happens to have. The takeaway for investors is that compute capacity, not model benchmarks, is the durable competitive moat in enterprise AI revenue today.
xAI's Colossus is currently running at roughly 11 percent utilization on its 550,000 Nvidia GPU footprint, with only around 60,000 chips actively in use, compared to roughly 43 to 46 percent at Meta and Google. The 89 percent gap is not idle hardware sitting on a shelf, it is an efficiency problem that includes server throughput design and the operational reality that research teams hold GPUs idle to preserve capacity for the next experiment. Even the strongest labs benchmark around 40 percent utilization, so the upside from operational maturity alone is substantial. If Colossus pushes utilization toward 50 to 60 percent, the incremental compute that becomes monetizable rivals the output of a top three hyperscaler. SpaceX has already stated an intent to build orbital data centers but the terrestrial data center business can take enterprise customers now. A meta-layer is also emerging in companies like Baseten, which sit on top of 18 or more data center providers and route inference workloads to whichever provider has the lowest cost and highest idle capacity, which directly converts utilization gaps into revenue. We view SpaceX's compute footprint as one of the most underappreciated revenue stories in the AI infrastructure stack. [Post-pod the Anthropic + xAI data center deal was announced so SpaceX is officially a hyperscaler.]
The framing of AI compute as the next oil is roughly directionally right but understates the supply problem, because oil supply is plentiful and well distributed while compute is structurally short and slow to bring online. Chip improvements, electricity production, and humanoid robotics could accelerate data center buildout, but the lead time on a meaningful capacity increase is still measured in years. The result is that enterprises will be forced into an in-house versus outsource decision on compute, exactly the way they decide on employees, with multi-year AWS or Google Cloud contracts locking current pricing but renewals at risk of doubling on a same-store basis. Lambda's thesis of shipping server stacks for local installation becomes economically rational the moment cloud compute pricing crosses the in-house breakeven (Lambda also has a massive AI cloud business). Apple is already pulling demand into the local model with AI-enabled Mac mini and Mac Studio, both of which are now back-ordered for OpenClaw workloads. The corporate signal is the same direction: Coinbase cut 14 percent of headcount this week citing AI, Block has done the same, and the work itself does not disappear, it converts from human calories into computer tokens. We expect compute cost per token to become an explicit line item in CFO planning by 2027, and the AI Platform companies (Anthropic, OpenAI, xAI, etc) that solve compute supply will capture an outsized share of the resulting spend.
The cleanest pre-IPO opportunity in AI is not Anthropic, OpenAI, or Databricks, all of which are excellent companies but fully valued ahead of their IPOs. Anthropic at a trillion dollar private mark with a likely IPO in the $1.2 to $1.3 trillion range offers roughly 30 percent upside before counterparty risk, secondary fees, and lockup, which is an inferior trade to simply buying the stock on day one in a brokerage account for free. The investable pre-IPO trade is in AI infrastructure ... in our opinion: AI data center neo-cloud providers, AI semiconductor companies focused on inference, and the routing and efficiency layer sitting on top. Baseten is a particularly interesting name, with $1 billion of forecast revenue this year, customers including Cursor, and a $4.7 billion secondary market valuation, which is small cap pricing for a business at the center of inference workload routing. Fireworks AI is another routing and efficiency play. Inference silicon names include Groq, Cerebras (IPO pending), and Positron. AI infrastructure focused on inference compute has potential to produce hundreds of billions of dollars of enterprise value. These businesses generally trade below $10 billion in private markets, are doing hundreds of millions of revenue, often have Nvidia on the cap table, and sit directly in the path of every AI dollar deployed by enterprise. The risk-adjusted return profile in this segment is, in our view, the most asymmetric setup in the private markets right now, with a credible path to 3 to 7 times capital before public listing, in our opinion.