When Kimi K3 landed, the headlines said Chinese open-weight models had caught the American labs, and the AI trade sold off from chipmakers to the labs themselves. Nobody stopped to ask whether cheaper and more profitable mean the same thing.
In this week's AI to ROI Big Story, Ray Rike and Peter Buchanan run the actual business math on both sides of the fight and find that neither side has the balance sheet to fight a sustained price war.
The setup is stark. Anthropic is projecting its first-ever quarterly operating profit of roughly $559 million in Q2, with an annualized revenue run rate near $47 billion, up from $9 billion at the end of last year. That profit disappears the moment it tries to match open-weight pricing. Gross margin is running around 40%, roughly 10 points below the internal forecast, against more than $350 billion in data center commitments coming due over the next three to five years. OpenAI's picture is even thinner: roughly $30 billion in ARR, a projected $14 billion operating loss this year, data center commitments approaching $1 trillion, and an advertising business off to a slow start that needs to reach $100 billion by the end of the decade to close the gap.
What the episode covers:
- Why the Chinese open weight labs are not the subsidized price killers the coverage assumed, with Z.ai's gross margin falling from 41% to roughly 15%, DeepSeek near break-even on about $500 million of revenue and already back in market after a $7 billion round, and Moonshot and MiniMax raising at rising valuations rather than running toward profit
- The open weight versus open source distinction that changes the entire economic model, since every new customer requires more chips, power, and data center capacity, and Z.ai's own numbers show roughly 49 to 50% gross margin on customer hosted deployments versus about 19% when they host and serve via API
- The price war math itself: frontier models cost roughly $6 to $8 per million output tokens to serve, Anthropic's $25 per million on Opus produces about a 70% gross margin, and repricing down to the $4 to $6 range where Meta's Muse Spark sits flips that margin from positive 70% to negative 65%
- Why DeepSeek cut prices on a low-end model and then, two weeks later, told customers to prepare for substantial increases across the line, particularly on API access
- Google as the structural outlier, with 83% growth in its cloud and AI segment, Gemini embedded across fifteen products with more than a billion users each, and the Apple Siri deal extending reach toward two billion devices
- Where Kimi K3 actually fits, including the caveats nobody is pricing in: weights released only last week, no published large-scale production deployments, two to three times the token consumption on complex tasks, and infrastructure requirements around a 72 GPU rack that costs millions to install and millions a year to run
- The geopolitical wildcard, with Washington weighing sanctions or outright bans on Chinese open-weight models and distillation-related IP exposure still unresolved
The metric that resets the argument: cost per completed task
Price per token is the easiest unit to measure and the wrong one to buy on. Ray and Peter walk through a frontier lab evaluation that assumed a fully loaded remediation cost of $17 per failed attempt, roughly 10 minutes of a human operator's time. Claude Opus completed the task about 90% of the time at roughly $2.56. Meta's Muse Spark, priced at a quarter of Opus on tokens, succeeded 75% of the time and landed above $5.80 per completed task. The list price was 75% lower, and the delivered cost was more than double. A fifteen-point reliability gap did all the work.
The formula Ray offers turns a squishy quality debate into something a CFO can actually evaluate: cost per attempt, plus failure probability times fully loaded remediation cost, divided by success rate.
What CFOs and GTM leaders should take away:
- Build cost per completed task into vendor evaluation and make vendors compete on that number rather than on a token price list
- Segment AI workloads by what a failure actually costs, since a wrong answer in a regulated process or a customer support interaction carries a very different price than the model delta suggests, and standardizing on one model across both workload types to save on price is the common mistake
- Treat orchestration as a cost lever, not just the model choice, since Cursor's internal testing found coordination across multiple models delivered comparable code quality at a fraction of the cost of a single large model
- Run your own evaluations instead of trusting public leaderboards, since LMSYS Chatbot Arena measures human preference rather than task completion and says nothing about your workload
- Watch the switching trap, because a cheap model that attracts heavy traffic today can reprice two or three times higher in a quarter once the vendor needs margin
- Do not overreact to the open weight scare by standardizing on the cheapest option, since compute, power, and people costs are accelerating, and vendor durability still belongs in the evaluation
Ray's read: on real-world performance and total cost of ownership, the closed-weight labs still hold the advantage, their prices keep coming down, and they have every incentive to avoid a price war. Peter's read: open weight economics favor whoever hosts the model more than whoever built it, which makes the hyperscalers the quiet winners regardless of how this resolves.
For the full analysis behind this week's big story, subscribe to the AI to ROI newsletter at: ai2roi.substack.com
See Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.
