IaaS Providers & Aggregators — inference providers (Together, Fireworks, Groq) and aggregators like OpenRouter remain a major commercial route for open-weight models. However, several model creators now monetize inference directly — DeepSeek, Mistral, Moonshot (Kimi) and MiniMax all operate paid APIs alongside their open releases. Chinese providers in particular have seen explosive API revenue growth in early 2026, with Moonshot and MiniMax now ranking #1 and #2 globally on OpenRouter by token volume. Meta remains the notable exception — Llama generates $0 in direct inference revenue for Meta.
VC/Investor subsidy ($12B) is purely the gap between what customers spend on AI and the published, calculated and estimated operating expense to provide that service. It excludes training compute capex, data centre construction, and GPU procurement — which are funded separately and dwarf customer revenue.
Strip away the VC subsidy and the Trad SaaS revenue that ends up as margin (not token consumption), and the picture becomes stark:
Even ignoring people costs, R&D, and SG&A — the core AI inference business loses money at the gross margin level. Every token served costs more than customers pay for it. The $2.15B of "margin" in the system belongs entirely to Trad SaaS vendors and hyperscaler cloud fees — not to the companies building the models.
Inference costs need to drop from ~$13B to ~$8.4B for a minimally viable 30% gross margin. This means cheaper silicon, better hardware utilisation, and model efficiency gains. If this happens, NVIDIA's pricing power collapses — and that's where most of the $13B ends up today.
Open models (Llama, DeepSeek, Qwen) via IaaS providers have a thin but positive gross margin on inference (est. 15-35%, vs negative for closed models). No IaaS provider has operating margin yet — the GM is consumed by people and ops. But they're 5-10x cheaper per token, the fastest-growing segment (1.2B Llama downloads, 61% of OpenRouter token volume), and trail frontier by just ~3 months. A shift of ~20-30% of workloads to open models reaches system breakeven — without any hardware cost reduction.
At current inference costs and model mix, token and subscription prices need to rise ~55% to achieve 30% gross margin. Even breakeven requires an 8% price increase. But prices have been falling ~80% year-on-year as providers undercut each other with VC-subsidised pricing. Raising prices risks accelerating the shift to open models — the very thing that threatens closed providers' relevance.
In practice, all three are happening simultaneously — but in opposing directions. Inference costs are falling (good), open model adoption is accelerating (bad for closed providers), and prices are falling not rising (terrible for margins). The current closed-model economics — where every token costs more to serve than customers pay — cannot survive without a fundamental structural shift.
This diagram only shows the customer revenue dollar flow. The AI industry's total investment is dramatically larger:
For every $1 customers spend on AI, investors spend $15+ building it. Training compute, data centre construction, and GPU procurement are funded by VC rounds, corporate capex, and sovereign wealth — not customer revenue.
47% of all AI spending flows to inference compute — the variable cost of serving models to customers. NVIDIA and the cloud providers capture this regardless of which model wins.
The AI industry earned $15.6B but spent $27.6B. Investors covered the $12B operating gap. OpenAI lost ~$6B, Anthropic ~$5.2B. This subsidises free tiers, price wars, and the talent arms race.
The only margin in the entire system comes from Trad SaaS vendors (Microsoft, Salesforce, ServiceNow) and hyperscaler cloud margins. Frontier providers and IaaS providers have zero operating margin — IaaS has a thin positive gross margin on inference, but it's entirely consumed by people and ops costs.