The Great AI Downgrade: Why DoorDash, Airbnb, and Siemens Are Moving AI Workloads to Chinese Open-Weight Models

Georgi Kuzmanov Senior Equity Analyst & Founder Columbia MSFE · 13+ years equity research

For two years the enterprise AI trade rested on a simple assumption: frontier intelligence is scarce, so whoever owns the best model — OpenAI, Anthropic — owns the pricing power, and whoever resells it at a markup rides along. That assumption is now being tested by the corporate procurement department, the least sentimental buyer in capitalism. The new evidence says the assumption is breaking.

The Catalyst

According to the Financial Times, a growing roster of large Western companies — DoorDash, Airbnb, and Siemens among them — is now routing artificial-intelligence workloads to Chinese models from DeepSeek, Z.ai, and Moonshot AI. The draw is twofold: dramatically lower prices, and the “open-weight” architecture of most Chinese systems, which hands customers the model parameters themselves — allowing fine-tuning on proprietary data and full control over where sensitive information is processed.

The anecdotes behind the FT story are concrete. DoorDash co-founder Andy Fang said in a recent post on X that the delivery platform is saving significant money by handing “lower-level” work to a Moonshot AI model. San Francisco automation startup Lindy has dropped Anthropic’s models entirely in favor of DeepSeek’s new V4 family; its founder Flo Crivello put the logic memorably to Rest of World: “You don’t need God to write your email.” And usage data from OpenRouter, the routing platform that has become the industry’s de facto scoreboard, show models from DeepSeek and Z.ai are now used more heavily than Anthropic’s Claude or OpenAI’s ChatGPT-family models across its corporate traffic.

The timing is not accidental. This is landing weeks after Z.ai shipped GLM-5.2 on June 16 — a 753-billion-parameter open-weight model with a one-million-token context window that beats OpenAI’s GPT-5.5 on several long-horizon coding benchmarks (62.1 vs. 58.6 on SWE-bench Pro; 74.4% vs. 72.6% on FrontierSWE, per published results) at roughly one-sixth the per-token cost. “Good enough” stopped being a euphemism for “worse” sometime this spring; for a large class of workloads it now simply means “cheaper.”

The Landscape

Map the players and an unusual feature of this fight jumps out: almost none of the combatants are directly investable. OpenAI and Anthropic are private, with the two firms reported at roughly $24 billion and $30 billion in annualized revenue this spring, as we covered in April. The Chinese challengers — DeepSeek, Z.ai (maker of the GLM family), and Moonshot AI (maker of Kimi) — are private too. So is OpenRouter, the traffic exchange keeping score. The stock market only sees this war through proxies:

• The adopters (public): DoorDash (DASH), Airbnb (ABNB), and Siemens (SIEGY) are the named switchers — for them, cheap inference is a cost line improving, not a business model threat.
• The exposed (public): Microsoft (MSFT) carries the OpenAI relationship, Copilot pricing, and its own Azure AI margins into this fight; Amazon (AMZN) and Alphabet (GOOGL) are Anthropic backers and hyperscale hosts; Meta (META) is the one US giant whose open-weight Llama line competes on the challengers’ turf, even as it pivots toward proprietary products.
• The arms dealers (public): Nvidia (NVDA), AMD (AMD), and Broadcom (AVGO) sell the compute that all of these models — American or Chinese, open or closed — consume. The VanEck Semiconductor ETF (SMH) is the basket expression.
• The Chinese proxies (public, imperfect): Alibaba (BABA) and Baidu (BIDU) run their own model families (Qwen, Ernie), and the KraneShares CSI China Internet ETF (KWEB) is the standard China-tech wrapper — but note carefully: DeepSeek, Z.ai, and Moonshot are not inside any of them.
• The software layer: the iShares Expanded Tech-Software ETF (IGV) holds the application companies that both pay for inference and resell it at a markup.

By the Numbers

Prices below are July 14, 2026 closes; year-to-date returns are measured from the December 31, 2025 close. Market caps and trailing P/E ratios are computed from latest filings.

Read that table as a story and it tells you the market has already picked a side, crudely. The sellers of compute are having a historic year (SMH +66.7%, AMD nearly tripling) while the buyers and resellers of closed-model intelligence are bleeding: software (IGV) is down 11.4% in a year when the S&P 500 is up 10.3%, and Microsoft — the most levered large-cap expression of premium closed-model economics — is off 20.4% and sits 30.7% below its 52-week high. What the market has not done is reward the Chinese complex for its labs’ success: KWEB is down 23.1% this year. That disconnect is the most interesting cell in the table, and we return to it below.

The Shift

The source story is about a handful of company anecdotes. The data underneath it describes something much bigger: a collapse in the market share of American closed models on neutral routing infrastructure, driven by a price gap so wide it functions as a different product category.

Three forces are compounding. The first is the bill shock. Corporate AI budgets set in the pilot era are colliding with agentic workloads that consume tokens by the billion. Ramp’s AI Index for June 2026, drawn from card and bill-pay data across 70,000+ firms, found the top 1% of AI spenders now paying about $7,450 per employee per month for AI services — against a median firm’s $11 — and growing that spend 14% month over month. At the extreme tail, Axios reported that one large company ran up roughly $500 million in a single month on Anthropic’s Claude after failing to cap employee licenses. Whether or not that anecdote is repeatable, every CFO who read it understood the direction of travel: usage is exploding, and per-token price is the only lever procurement actually controls.

The second force is that the quality gap has narrowed below the threshold where most workloads notice. GLM-5.2’s published coding-benchmark wins over GPT-5.5 matter less for bragging rights than for permission: they give enterprise architects cover to route the 80% of volume that is classification, extraction, summarization, and routine code to a model that costs one-sixth as much, reserving frontier models for the hard 20%. This tiered-inference pattern — frontier model as escalation layer, open-weight model as workhorse — is precisely what Ramp observes among its heaviest spenders, who deliberately mix frontier and cheap open models to avoid lock-in.

The third force is trust, and it cuts in a direction few predicted in 2024. When the Commerce Department abruptly suspended foreign access to Anthropic’s Fable 5 and Mythos 5 models on June 14 — the episode grew out of the security anxieties we chronicled during April’s Mythos scare — overseas customers experienced something new: an American frontier model being switched off by government order. The controls were lifted on June 30, but the lesson stuck. Cohere’s CEO Aidan Gomez told the FT the ban was the clearest demonstration yet of the danger of depending on a single provider for critical AI workloads. Open weights are, among other things, an insurance policy: parameters you host cannot be revoked by anyone’s export-control directive.

Winners & Losers

Winners

• Nvidia (NVDA), AMD (AMD), Broadcom (AVGO): cheaper tokens mean more tokens — the Jevons dynamic. Inference demand is price-elastic, and every model in this fight, Chinese or American, runs on their silicon. NVDA’s $5.1 trillion market cap — a threshold it crossed the week of July 6 — is the market pricing exactly this.
• The adopters — DoorDash (DASH), Airbnb (ABNB), Siemens (SIEGY): heavy AI consumers with zero model-vendor loyalty. A 25x cut in unit inference cost drops straight into opex. DASH, down 34% from its high partly on AI-spend fears, is the clearest “cost-relief” candidate among the named names.
• Alphabet (GOOGL): the only closed-model vendor that owns its silicon end to end. TPU vertical integration lets Gemini chase the price war profitably rather than subsidize it — one reason the stock is +14.9% YTD while its Redmond rival is −20.4%.
• Meta (META): ambiguous but net-positive — commoditized intelligence is the world Llama was built for, and Meta consumes far more inference than it sells; the caveat is its own pivot toward proprietary models blurs the strategy.

Losers

• Microsoft (MSFT): the most exposed public expression of closed-model economics — premium Copilot seat pricing, Azure margins on OpenAI traffic, and a reported internal pullback in AI tooling spend. The 20% YTD decline says the market sees it too.
• Application software (IGV): companies whose product is a thin margin over a frontier API are watching that margin get repriced by an $0.87-per-million-token competitor. IGV’s −11.4% YTD against a +10.3% S&P 500 is a 21-point relative-strength indictment.
• Amazon (AMZN): partially — its multibillion-dollar Anthropic stake absorbs the valuation pressure of a price war, though AWS wins whenever inference volume grows, whoever’s model runs it.
• Baidu (BIDU): a reminder that “Chinese AI wins” does not mean “Chinese AI stocks win.” Ernie is losing the domestic developer race to DeepSeek and the open-weight upstarts, and Baidu’s trailing earnings have collapsed to the point where its P/E is not meaningful.

Risks & Counterpoints

Two further caveats. First, the benchmark wins that anchor the GLM-5.2 story are vendor-published and narrow (long-horizon coding); independent evaluations have a history of sanding down launch-day claims. Second, the open-weight advantage on data control is partly rhetorical — most enterprises consume “open” Chinese models through hosted APIs anyway, where the sovereignty benefit evaporates and the security questions return.

The Investment Angle

None of what follows is investment advice; it is a map of how the theme is expressible in public markets.

The infrastructure expression (cleanest): if cheap tokens beget exponentially more tokens, compute demand is the least thesis-dependent winner. That is SMH at the basket level, or NVDA/AMD/AVGO in single names — with the caveat that after a +66.7% YTD run in SMH and AMD at 178x trailing earnings, the entry price now embeds a lot of Jevons.

The cost-relief expression (least crowded): heavy AI-consuming platforms with no model to defend — DASH (−17.1% YTD, 34% below its high) is the named adopter where falling inference costs most directly repair a bruised margin narrative; ABNB and SIEGY carry the same logic with less drawdown to recover.

The relative-value expression: the pair implied by the whole piece is long compute / short closed-model resale — in liquid terms, overweight SMH against IGV, or GOOGL (owns its silicon, +14.9% YTD) against MSFT (rents its frontier, −20.4%). Note that this pair has already worked violently this year; it is a momentum position now, not a discovery.

What we would not do is buy KWEB or BABA/BIDU as a “Chinese AI” play on this news. The labs winning the token war — DeepSeek, Z.ai, Moonshot — are private, give their weights away, and monetize thinly by design (recall Vercel: 17% of tokens, ~1% of revenue). KWEB’s −23.1% YTD is the market telling you the listed Chinese internet complex does not capture this value; the trade there needs a monetization catalyst that does not yet exist.

The AlphaEdge Take

The FT story reads like a procurement anecdote; it is actually a regime marker. The era in which frontier-model quality alone could command a 30x price premium across the entire workload stack is over — killed not by a single Chinese model but by the combination of a 97% price discount, benchmark parity on routine work, and a June export-control episode that taught every non-US buyer what single-vendor dependency costs. Intelligence is bifurcating into a commodity tier and a frontier tier, and the commodity tier just went to the low-cost producer, which happens to publish its weights and happens to be Chinese.

What would change our mind: a US regulatory strike against domestic use of Chinese models (the Airbnb congressional scrutiny is the tell to watch), a security incident that makes open Chinese weights radioactive to general counsels, or evidence that frontier-tier pricing is eroding too — if OpenAI or Anthropic ever cuts top-model prices toward Chinese levels, the profit-pool assumption underneath a lot of AI-adjacent equity value goes with it.

For investors, the discipline is to separate the three claims the market keeps conflating: Chinese labs winning token share (true, measured), Chinese stocks winning value (false so far — KWEB −23.1%), and US AI capex breaking (unsupported — the compute complex is still the toll road under both outcomes). Own the layer where the value provably pools, rent the cost-relief stories while they are cheap, and treat the model layer itself — on both sides of the Pacific — as the place where margins go to be competed away.

Bottom line: the AI price war has turned model intelligence into a commodity faster than the market expected — own the compute toll road and the cost-relief beneficiaries, not the pricing power that is being competed away.

Georgi Kuzmanov

Senior Equity Analyst & Founder at AlphaEdge. Columbia University MSFE (2011–2013). Covering equities, macro, and geopolitics for serious investors.