Frontier labs as businesses.
Revenue, valuation, deployment posture, and risk. Numbers are repo-backed and marked by source type and confidence.
Zhipu AI / Knowledge Atlas
Sovereign enterprise and government deployment demand is the monetization wedge. GLM should be analyzed less like a consumer app and more like private-cloud AI infrastructure. Update, H1 2026 interim results: that wedge is being abandoned in front of us. Revenue grew 399.7% while consolidated gross margin halved from 50.0% to 26.4%, because the growth came from cloud API at 24.6% margin and the private-deployment business actually shrank 20.5%. R&D ran at roughly 8.5x gross profit. Zhipu is the only listed pure-play frontier-model lab, so this is the one place the model-layer P&L is audited rather than briefed - and what it shows is a lab growing by selling more of a thinner thing, with research spend, not compute, as the line gross profit cannot cover.
iFlytek (科大讯飞)
iFlytek is the control group. It is the oldest AI-native listed company in this file, it built a two-decade speech-recognition moat, and it has published audited financials straight through the LLM transition - which makes it the only place to see what that transition did to a real P&L over time. The answer is: nothing, to the margin. Gross margin sat between 40.69% and 42.59% every year from FY2021 to TTM June 2026, with the Spark LLM launching in the middle of that window. That is the strongest single argument that Zhipu’s 26.4% and MiniMax’s 17.9% are the economics of a young pure-play scaling API volume, not the economics of serving models. What the LLM era did hit is growth and the underlying loss: H1 2026 revenue grew 6.52%, the slowest here, while the non-GAAP loss widened 65-85% even as the reported loss narrowed. The competitive read is the sharpest datapoint iFlytek offers - rivals did not out-build the speech moat, they routed around it. A team fine-tuning an open-weight foundation for a few million yuan reaches roughly 80% of the capability, and ByteDance’s Doubao input method took 50 million downloads in a week by competing on semantic understanding instead of recognition accuracy. The wall still stands; the enemy stopped crossing the moat.
SenseTime (商汤)
SenseTime is the counter-example that keeps the China-lab read honest. It is the pivot case - a computer-vision company that rebuilt itself around generative AI, which is now 79.9% of revenue - and its H1 2026 goes the opposite way to Zhipu on every line that matters: gross margin UP 2.9 points to 41.4% while Zhipu halved to 26.4%, gross profit growing faster than revenue, and the first IFRS profit since the 2021 listing. Two cautions keep it from settling the argument. The profit is substantially fair-value gains on strategic investments in AI ecosystem companies; the non-IFRS adjusted result is still a RMB 390M loss, and both descriptions of the half are accurate. And "generative AI revenue" here includes selling compute through its Token Factory, which is a different business from selling model access and carries different economics. What SenseTime does prove is that 41.4% is achievable by a company that is 80% generative AI by revenue - so the 18-26% at the young pure-plays is a stage, or a business-model choice, not a law. It also publishes the one number nobody else does: 2.4 trillion tokens served daily in July 2026, roughly 22x YoY, against 48,000 petaflops deployed.
DeepSeek
DeepSeek is the canonical sparse open-weight economics case: architectural efficiency and open distribution can reset price expectations even when revenue is opaque. Update, 2026: revenue is becoming less opaque and the efficiency claim is now testable. A $400-500M run rate is reported alongside 70-80% gross margins on V4 flagship access - roughly three times Zhipu’s audited 26.4% while pricing below it. A STAR Market listing would put that claim under audit, which makes DeepSeek the single most informative pending disclosure in this file.
Moonshot AI
Moonshot is the test of whether a China-native consumer AI app can convert long-context pull into durable subscription or API economics.
MiniMax
MiniMax is the app-native China lab pattern: global consumer and creative tools can grow fast, but losses show that app distribution does not remove frontier burn. Update, H1 2026 results: the app-native framing is now half wrong. Revenue grew 283.1% to $116.6M, more than all of FY2025, but the growth was enterprise API - Open Platform up 703.1% to 63.4% of revenue - while consumer products merely doubled. Selling expense FELL 17.9%. So distribution is not the engine; developers are. The burn point holds and hardens: gross margin 17.9%, net loss $358.0M, R&D 14.3x gross profit.
Anthropic
Anthropic is the code-agent and enterprise reliability monetization case, with compute procurement as the visible margin constraint.
OpenAI
OpenAI is the consumer distribution and frontier API scale case. ChatGPT is the asset; the model alone is not the moat.
xAI
xAI is the physical-scale compute-first AGI case. It leverages Memphis Colossus scale-out and deep integration inside the Tesla/SpaceX/X ecosystem to accelerate real-world reinforcement learning. Update, February 2026: xAI stopped being a standalone entity. SpaceX acquired it all-stock at $250B and now reports it as one of three segments, which makes it the only US frontier lab whose numbers arrive through an earnings call - $2.56B of AI segment revenue in Q2 2026, up 247%. Two cautions. The perimeter is wider than the lab: that figure includes Grok subscriptions. And the same quarter carried $18.4B of company capex against it. The open question is whether an AI business inside a capital-intensive parent ever gets judged on its own return, or is simply carried.