The AI price war: OpenAI and Anthropic take aim at China - Capacity Global

OpenAI and Anthropic slash prices as Chinese rivals undercut them by up to 9x, reshaping enterprise AI and data centre demand. OpenAI has cut the cost of its lightweight GPT-5.6 Luna model by 80%, taking API pricing down to $0.20 per million input tokens and $1.20 per million output tokens, while trimming its mid-tier Terra model by a fifth. Anthropic, whose Claude models have become the default choice for enterprise coding and agentic workflows, has responded with Claude Opus 5, pitched by the company as delivering frontier-level performance at roughly half the cost of its previous flagship. Sam Altman was characteristically blunt about the motivation on social media, noting that GPT-5.6 Sol already undercut Anthropic’s Fable 5 and that OpenAI would be “happy to deliver at one-quarter of the price” if that is what competition demands. The reason neither company can afford to sit still is sitting in Beijing, Hangzhou and Singapore. Chinese labs including DeepSeek, Moonshot’s Kimi and Zhipu’s GLM have closed the capability gap enough that, for a large share of everyday enterprise workloads, buyers simply cannot justify the premium anymore. Running an equivalent task through Zhipu’s GLM model has been benchmarked at roughly $544, against $4,811 for the same job on Claude, a near ninefold gap that finance departments have started putting into spreadsheets rather than ignoring. Why the price cuts matter beyond the API bill For technology and data centre leaders, this is not simply a story about cheaper chatbots. It is a signal about where compute demand actually lands, and how quickly it can move. Anthropic’s own Claude Opus 5 launch was framed around “frontier intelligence . . . at half the price” of its predecessor, language that only makes sense if the company expects volume to compensate for thinner margins per token. That is the same logic hyperscalers have used to justify years of capacity expansion: cheaper access drives usage, and usage drives infrastructure build. If the price war succeeds in pulling enterprise workloads back from Chinese platforms, US hyperscale and colocation operators should see sustained, possibly accelerating, demand for GPU-dense capacity. If it fails, and cost-conscious customers keep migrating to Chinese-hosted or open-weight alternatives run on sovereign infrastructure, the demand curve for Western-aligned data centres looks very different. There is also a geopolitical layer that data centre operators cannot ignore. Washington’s export controls have already reshaped who can access which models: Anthropic was reportedly given a 90-minute window to withdraw Fable 5 access for foreign nationals, while OpenAI’s flagship GPT-5.6 tier has been gated to a shortlist of roughly 20 approved partners. Dario Amodei has argued for years that restricting Chinese labs’ access to advanced chips is the correct policy response, writing that DeepSeek’s progress proved controls were “a whack-a-mole model” only in the sense that Chinese firms would need ever more power and workaround engineering to match Western output. Those building or leasing capacity in jurisdictions caught between the two regulatory regimes, the Gulf, Southeast Asia and parts of Europe among them, are now pricing that political risk alongside power and connectivity costs. Nobody has yet mapped the price war onto actual colocation and hyperscale leasing patterns. If enterprise AI spend is moving toward usage-based billing rather than flat subscriptions, as both OpenAI and Anthropic have been doing, the revenue volatility that creates for AI labs should eventually show up in how aggressively they commit to long-term power purchase agreements and build-to-suit contracts. That is a story about credit risk and counterparty profile as much as it is about chip pricing, and it deserves the same treatment as the recent analysis on how AI-driven demand is straining power grids and cooling supply chains . Secondly, the sovereignty angle needs more scrutiny. Chinese labs’ hosted APIs route data through servers subject to Chinese national intelligence law, a point Western vendors have leaned on hard in their own marketing. But sovereign and regional buyers, particularly in the Gulf, are already weighing data residency, latency and political alignment against raw price, a tension explored in coverage of how the Iran-US conflict exposed data centre supply chain risk in the region. As DeepSeek pursues its own IPO ambitions, reportedly targeting a $74 billion valuation ahead of a China listing , the question of which sovereign and enterprise buyers it can realistically serve outside China becomes a live commercial issue, not a theoretical one. Third, there has been little scrutiny of what cheaper AI does to the underlying hardware and power squeeze that has dominated data centre coverage all year. JLL’s Andrew Batson has described AI training facilities as demanding ten times the power density of traditional halls and commanding “60% lease rate premiums” over conventional data centres. A price war at the model layer does not reduce that physical constraint; if anything, cheaper tokens should increase inference volume and push more workloads towards the power-hungry infrastructure already covered in reporting on AI’s toll on shortages of memory chips and cooling capacity . What it means for future deals For those negotiating capacity agreements with AI labs and neoclouds over the next 12 to 18 months, the price war changes the underwriting conversation. Frontier labs racing towards trillion-dollar IPO valuations, Anthropic among them, need to show investors that falling per-token prices are offset by rising volume rather than eroding margin. That makes them more likely to seek flexible, scalable capacity commitments rather than the fixed, decade-long leases hyperscalers have traditionally favoured, and more likely to treat power and cooling availability, not model quality, as the binding constraint on growth. Buyers of colocation and connectivity should expect AI-lab tenants to negotiate harder on ramp schedules and exit terms even as they commit to larger footprints, and should watch enterprise adoption data as closely as they watch chip shipment figures. The labs that win the pricing fight will need somewhere to run the workloads it generates; the infrastructure side of that equation, not the API price list, is where the next round of deal-making will actually be decided. Related stories Trump renews push for AI giants to foot data centre power bill VIRTUS Data Centres appoints Adam Eaton as new CEO to drive digital growth What is behind-the-meter power and how is it reshaping the data centre industry? Datacloud USA 01 September 2026 Register now See event details Nadine Hawkins Director of Content and Insights Share Share on LinkedIn Share on Twitter Copy Link URL Top Articles Top articles --> data centre Texas hits pause on its own AI boom: Inside the 300-project grid audit 17 August 2026 5 minutes Data Centres 8 sessions to watch at Metro Connect Fall x Datacloud USA 2026 17 August 2026 4 minutes News Nvidia halves Ohio OpenAI backstop to under $120bn: Why it happened, and why it matters 17 August 2026 4 minutes Data Centres The AI price war: OpenAI and Anthropic take aim at China 17 August 2026 6 minutes News Lenders start pricing community opposition into data centre financing as $130bn in projects stall 14 August 2026 3 minutes