The financial projections shared by OpenAI and Anthropic with investors paint two starkly different pictures of how to reach superintelligence. OpenAI's roadmap shows a dramatic spike in training costs between 2027 and 2028—doubling from $60 billion to $125 billion annually—suggesting a leap-year bet on raw computational scale to unlock new capabilities. Anthropic's trajectory climbs more modestly, assuming roughly linear cost growth around $30 billion by 2029. This gap hints at fundamentally different engineering philosophies: OpenAI betting on brute-force scaling, Anthropic wagering that efficiency breakthroughs will reduce the need for exponential capital.
Neither company expects near-term profitability if training costs remain as projected. OpenAI's models show training consuming roughly 100% of revenue through 2027, only dipping below that ratio by 2029—meaning the IPO most observers expect within two years will land on a company burning tens of billions annually. Anthropic stays more conservative, keeping training costs under 100% of projected revenue throughout. The timing matters: public markets punish sustained losses differently than private venture capital does, yet both firms are implicitly wagering that 2028 brings either a breakthrough in efficiency, a plateau in capability gains, or explosive revenue growth that outpaces cost escalation.
The real vulnerability lies in inference costs, which currently consume roughly half of OpenAI's revenue. While hardware improvements and algorithmic gains should reduce per-token costs, demand dynamics may work against them. As AI adoption scales from billions of users to billions of concurrent users deploying complex agentic workflows, token consumption won't grow linearly—it will explode combinatorially. Sub-agents spawning sub-agents, code context expanding tenfold, reasoning loops running in parallel: these aren't minor implementation details, they're inherent to the next generation of capability. Whether falling per-unit costs can outpace exponential growth in total tokens consumed remains the hidden variable neither company can fully model.
Beyond raw economics, Anthropic's decision not to release Claude Mythos—despite leaked evidence it can identify critical OS vulnerabilities at sub-$2,000 cost—signals a maturity OpenAI has struggled to demonstrate. Project Glasswing, mobilizing Apple, Google, Microsoft, Amazon, and others to patch systems before widespread access to a weaponizable vulnerability, is governance under pressure. It's also strategic cover: Anthropic gets to be the responsible actor, the company that safeguards first and scales second. That posture matters as Congress grapples with AI regulation and the venture ecosystem looks for partners it can defend to regulators. OpenAI, meanwhile, pivots toward enterprise consulting and acquired a Twitter tech show, fighting what the founders clearly perceive as a reputation crisis in the Silicon Valley echo chamber. One company is optimizing for capability and speed; the other for trustworthiness and stakeholder alignment. Neither strategy is obviously wrong, but only one can dominate if capital becomes constrained.