THE $800B COMPUTE TRAP
Ben Lorica and Evangelos Simoudis work through Anthropic's S-1 filing, weighing its striking growth and improving unit economics against compute commitments, falling prices from open-weight models, and dependence on coding. They then examine SpaceX's emerging neocloud business and close with a debate over proposed frontier AI regulation, from self-improvement limits to independent testing and incident disclosure.
Anthropic’s S-1: growth that is hard to fake
The headline numbers in Anthropic’s S-1 are, by the consensus of this conversation, extraordinary. Revenue growth is the first thing that stands out, and Evangelos Simoudis argues it is hard to fake. The company also reports improved unit economics: revenue per dollar of compute spend reportedly moved from about 40 cents to about 2.28 dollars. Both participants caution that the gross margin figure needs scrutiny, especially what is included in cost of revenue, and that the numbers only run through the first half of 2026. Open-weight models have not yet visibly pressured pricing in enterprise, though enterprises are experimenting with smaller proprietary models to lower costs, and some are exploring open-weight models, including Chinese ones, for on-premise deployment where data stays in-house.
Cash burn looks less alarming than expected. The company has raised about $141.1 billion and holds about $131.7 billion in cash and short-term investments. Simoudis credits a deliberate approach to cash management, including pushing some expenses into warrants and debt. Rising interest rates could eventually make that approach costly, but near-term liquidity is not the concern.
Customer breadth and revenue concentration
Customer concentration has improved sharply. No single customer exceeds 10% of revenue, the top five share fell from 36% to 21%, and customers spending at least $100K grew from about 950 to nearly 12,000. The residual 4 to 8% of revenue flowing through cloud hyperscalers is a mixed signal, since those same hyperscalers compete with Anthropic. Two questions frame the risk: whether these customers keep their spending as they shift toward smaller or open-weight models, and whether Anthropic’s consulting arm can push deeper into the market. Simoudis notes that Meta and Microsoft, previously heavy users, have started pulling back on Claude, which he flags as a concern.
Revenue per employee rose from about $1.4 million to $6.6 million over twelve months. Simoudis sees two signals in that figure. It shows conventional enterprises what an AI-first company can achieve, and it raises the bar for SaaS incumbents such as ServiceNow, Salesforce, and Workday, and for traditional software firms like SAP, which will have to rethink how they improve their own per-employee economics.
Coding dependence and the SaaS question
The business is heavily concentrated in software development, with some estimates placing coding at 70 to 80% of revenue. Ben Lorica points to a familiar venture adage: early customers are rarely representative of the broader market. Coding was a remarkable product-market fit, but success in software does not automatically transfer to legal, finance, or insurance operations. Software is unusually amenable to AI, and the curves seen there may not repeat elsewhere. Some signs of product life beyond raw API usage, such as the shift toward Claude Code, are encouraging, but both speakers warn against overselling them.
The SaaS apocalypse narrative has been overhyped, though worries have not disappeared. Quarterly results from ServiceNow and Salesforce suggest investors still question the category. Simoudis argues that many companies are re-examining how much of a packaged application they actually use, now that building software internally is faster. Yet usage data often contradicts public claims of departure: people posting about leaving a platform are frequently still using it heavily. More decisively, SaaS vendors hold large volumes of customer data and are beginning to charge for access to it even when customers do not use the application’s business logic. That makes a wholesale collapse unlikely, even as AI tools exert real pressure.
Compute commitments and the training treadmill
Anthropic has reportedly committed about $800 billion in compute over the next ten years. Lorica notes that the company avoids the word “trillion,” and the figure still deserves scrutiny. Simoudis points out that it is unclear how much of the commitment data center developers can actually deliver. Roughly 60 to 65% reportedly cannot be cancelled, which is a meaningful but partial silver lining.
The cost structure is reinforced by a training treadmill. Frontier labs must keep releasing models to remain relevant, and each release carries heavy training costs. Simoudis argues that enterprises are not switching models as quickly as anticipated. Enterprises say that if no new model arrived for six to twelve months, they would still have more than enough to do. As experimentation gives way to deployment, model switching will slow, and the labs will have to find a different business model.
Prices are falling, driven by Chinese labs releasing open-weight models at industrial scale and by Western open-weight efforts gathering momentum. Lorica’s analogy of a water skier behind a speedboat captures the dynamic. Open-weight models can approach frontier performance as long as the proprietary models keep advancing, but they cannot overtake them. Falling prices pressure the growth narrative that justifies the valuation.
Switching costs and governance
Enterprise adoption is a major factor in durability. During experimentation, teams switch freely across models and tools. Once they move into deployment, switching becomes costly, and consulting partnerships may lock in those relationships. Lorica notes that an emerging wave of shared memory and shared context tooling may reduce this lock-in, since losing context is a common frustration when moving between coding tools.
Governance raises concerns too. The seven co-founders hold 50.1% of the votes, so public shareholders buy economics without control, a departure from many past IPOs. The company’s dilution profile adds to this concern. Milestone-linked stock issuance and share commitments to philanthropy could dilute shareholders significantly. Simoudis is not opposed to philanthropy but notes that it still matters from a shareholder perspective. Strategic partnerships with Amazon and Microsoft, which are also rivals, add another layer of tension.
SpaceX’s emerging neocloud business
SpaceX’s earnings suggest it is becoming a neocloud. Its AI segment reportedly contributed $431 million in operating profit while spending $16.4 billion, and sales and marketing rose 26-fold while revenue rose 14-fold. The company has three conditions to meet for this to work. The first is continued access to cheap capital. Elon Musk has argued that a bifurcation is coming between neoclouds that can access cheap capital and those that cannot, and that pure-play neoclouds will struggle. Simoudis agrees but notes that SpaceX is spreading its risk across many businesses, including the Spectrum acquisition and a possible car company. Each of those ventures is capital intensive, though. A telco can require $30 to $50 billion, and Terafab reportedly needs $100 billion.
The second condition is a base of solvent, long-term enterprise customers. Current neocloud customers include Anthropic and Reflection AI, neither of them profitable. Compute is a commodity, so premium pricing is unlikely. Simoudis believes SpaceX entered the neocloud business by accident: it built capacity around Colossus to train its own models, then found it could sell the surplus. Meta is now talking about similar capacity. The third condition is that spot prices for neocloud compute keep rising. Demand currently exceeds supply, but Simoudis expects a more balanced market by the end of next year. Lorica predicts that next-generation AI labs, which bought compute in advance to lock it in, may sell capacity to show growth before their next funding round, and that investors may be disappointed by what is actually delivered.
Regulating frontier AI labs
With midterm elections approaching and AI regulation likely to become a litmus test for the 2028 Democratic presidential nomination, several proposals are circulating. The first group targets self-improvement. Proposals to pause recursive self-improvement, or to use only vetted models to develop successors, both run into the same problem: no one has defined the term precisely. Using AI to write code might count, or it might not. Enforcing a restriction on using AI to work on the next generation of models would be nearly impossible, and no vetting process currently exists. Simoudis argues that the people designing these rules are typically behind the technology, and that whoever does the vetting, and how they were chosen, matters more than the proposal itself. A proposal to wait three to nine months before using newly released models for internal R&D is seen as impractical, since a model’s public safety record says little about its internal use. Coordinated slowdowns with competitors are widely viewed as a prisoner’s dilemma that no one believes can succeed.
Testing proposals run into capacity limits. Anthropic already gives employee-level access to certain agencies, and Simoudis argues that government agencies will struggle to hire qualified evaluators because of poor compensation and weaker internal tools. Private testers modeled on credit rating agencies have a precedent for capture, as the 2008 crisis showed when banks paid the agencies rating their securities. Underwriters Laboratories and European equivalents, which have long traditions in test engineering, offer a more promising template, though their attempts to extend beyond consumer devices have been less successful. Sharing models with international safety institutes is unlikely under the current administration. Financial-style stress testing for AI models is seen as more plausible.
On security, locking down training environments has gained traction. OpenAI reportedly now trains models inside a sandboxed copy of the internet, which Lorica describes as a larger sandbox that still allows agents to act freely. Simoudis argues that mandatory disclosure of security incidents would be a useful first step, since the public is only receiving piecemeal information about agent breakouts and may not know the full extent of problems. Whistleblower protections and rewards for internal escalation round out the list, though Simoudis cautions that the recent resignation of an Anthropic employee generated more noise than substance. Anyone raising alarms should be serious about what they are causing, given how anxious employees, consumers, and the public already are.