OpenAI

You say OpenAI paid Microsoft $17.2 billion in 2025. OpenAI's revenue was $13.07 billion. So OpenAI paid its largest investor more money than it made in revenue. How is this a business and not a money-laundering operation in slow motion?

The $17.2B includes both actual compute costs (R&D $10.59B + CoR $6.047B) and is partially offset by Microsoft's equity stake appreciation. The compute is real — model training and inference for 900M weekly users requires that infrastructure. However, the circularity is undeniable: Microsoft receives more cash from OpenAI than OpenAI generates in revenue, while simultaneously holding a $135B equity stake. The terminal question is who funds the next round when current investors are already owed more than current revenue.

OpenAI's audited 2025 operating loss is $20.92 billion. Sacra projects $27 billion in cash burn in 2026 and $63 billion in 2027. The company is targeting profitability by 2030. What has to be simultaneously true — inference costs, revenue growth, pricing power — for that path to exist?

Profitability by 2030 requires roughly: (a) revenue reaching $80-100B (possible if $24B ARR in 2026 grows 3x/year for 2 more years); (b) inference costs falling by 60-70% from 2025 levels (historical compute-cost curve suggests plausible but not certain); (c) R&D spending as % of revenue declining sharply (requires model capability maturation). Each assumption needs to hold simultaneously — and any one competitor breakthrough (open-weight models, Anthropic's cost advantage) breaks the pricing-power arm of the thesis.

The Stargate Project: $500 billion announced by Trump, OpenAI, SoftBank, Oracle. Multiple reports say it has hired no staff and no data centers are being actively developed. Where is the compute if not from Stargate?

OpenAI is relying on existing Azure capacity (committed $250B to Microsoft), CoreWeave ($11.9B deal), and bilateral Oracle capacity (4.5 GW deal replacing the Stargate JV). The CFO explicitly said at Davos that OpenAI is "working with partners to protect its balance sheet." Stargate was always partly a political announcement — the operational compute is under separate bilateral agreements, and those are real. The risk is that the fragmented bilateral approach is more expensive than an owned-and-operated data center would have been.

Anthropic just passed OpenAI on run-rate revenue at $30-45 billion ARR versus OpenAI's $24 billion. Anthropic spends 4x less on training per dollar of revenue. What does OpenAI's scale spending buy, and why does it matter if the cheaper lab just won on revenue?

OpenAI's counterargument is that frontier model capability drives enterprise adoption and consumer brand that trickles into all revenue tiers. The GPT-5 → o3 architecture may have a capability moat that justifies the training premium. However, the sourced data makes this uncomfortable: if Anthropic's more efficient approach generates more revenue, the presumed quality-spend-revenue correlation is broken. This is the most under-discussed risk in OpenAI's investment case.

Amazon's $35 billion of its $50 billion investment is contingent on OpenAI completing an IPO or reaching AGI. That's 70% of Amazon's commitment. What happens to OpenAI's balance sheet and its Stargate-replacement compute access if the IPO is delayed past 2028?

The contingent structure means OpenAI needs either a successful public offering or an AGI declaration to unlock $35B of Amazon's capital. If IPO is delayed — market conditions, regulatory challenge, or failure to demonstrate a credible path to profitability — that $35B reverts to a commitment-without-obligation status. Amazon's compute relationship (AWS as secondary cloud) may remain intact commercially, but the equity capital disappears. With projected $63B burn in 2027, that gap matters.