For compute buyers, gross margin is mostly an accounting-basis question. Nvidia's Aug 26 results (FY2027 Q2 8-K, unaudited) showed 75.0%; MiniMax's interim (Aug 26, unaudited) showed 17.9%, a gap of 57 percentage points. DeepSeek looks reversed: The Information reported (Aug 26, single source, unconfirmed by the company, cost basis undisclosed) seven-month revenue of RMB 475M at a 44.6% gross margin, implying cost of revenue of just RMB 260M, while the same piece put compute spending at RMB 11B, a 42x gap. All are unaudited interim figures; read only the direction.
Investors: before reading any model company's gross margin, get its basis (rented or bought compute, in cost of revenue or not, training amortization in or out).
Founders: do not raise on a gross margin that excludes training amortization; diligence will see through it.
Falsifier: if audited figures in DeepSeek's STAR Market prospectus roughly match these two gross margins, this entry is void; the prospectus basis governs.
Disclosure: written and verified with Claude tools; this entry discusses Anthropic, so discount favorable conclusions; FutureX is not on DeepSeek's or MiniMax's public shareholder lists (full lists go only to qualified investors).
This diary does not constitute promotion of, or an offer to subscribe for, any fund product.