Token costs fell 75 percent. Zhipu took the open weight lead. The data center math survives.

Per million token costs fell from roughly $10 to roughly $2.50 in a year per Ramp. Zhipu's GLM 5.2 dropped MIT licensed on June 13, sitting a point off Anthropic Opus 4.8 on FrontierSWE at a fifth of the cost.

Two CNBC stories landed on Friday June 26 that reshape the floor under the AI data center build. The first said the era of “tokenmaxxing” is over. The second said China’s open weight ceiling just rose to within a percentage point of Anthropic’s flagship. Read together, they reprice the inference layer underneath every hyperscaler capex print. The materials read survives the reshuffle.

What’s happening

Ramp’s enterprise spending data, as reported by CNBC on June 26, shows the average cost per million tokens across major model providers fell from roughly $10 to roughly $2.50 in a single year. Uber capped employee spend at $1,500 a month per AI coding tool after burning its 2026 AI budget in roughly four months. Lindy CEO Flo Crivello moved 100 percent of his agentic-AI traffic off Anthropic Claude onto DeepSeek hosted on US soil, telling CNBC the cost curve “crashed to the ground” and that the switch will save the roughly 25-person company millions of dollars within months.

Same day, CNBC’s second piece said Zhipu’s GLM 5.2, released open weight under an MIT license on June 13, scores 51 on the Artificial Analysis Intelligence Index v4.1. That places it ahead of MiniMax M3 (44), DeepSeek V4 Pro (44) and Kimi K2.6 (43) on the open weight ladder. On the closely watched FrontierSWE benchmark, GLM 5.2 trails Anthropic Opus 4.8 by 0.7 points (74.4 percent versus 75.1 percent) at roughly a fifth of the inference cost.

Both Anthropic and OpenAI filed confidential S-1 prospectuses in early June. Anthropic filed June 1 at a reported $965 billion valuation, the level set by its Series H. OpenAI followed June 8 at the $852 billion post-money reference from its March 2026 round, with underwriters reportedly working toward a near-trillion-dollar IPO range. The Microsoft Build 2026 keynote on June 2 unveiled MAI-Code-1-Flash, a 5-billion-parameter in-house coding model that Microsoft says outperforms Claude Haiku 4.5 by 16 points on SWE-Bench Pro and solves comparable problems with up to 60 percent fewer tokens on SWE-Bench Verified.

Brazil angle

Brazil never got to “tokenmaxxx.” Banks, insurers and retailers were already squeezed by FX and credit costs from the start of the AI cycle, so cost-per-inference was the operating constraint from day one. The shift the United States is now living through is the cycle finally arriving where Brazilian enterprise already lived. The open weight ceiling rising matters more here than any Anthropic price cut that has not arrived. GLM 5.2 and Kimi K2.6 are deployable today on locally hosted racks, exactly the workload profile that Brazil’s first-phase data center geography was sized for.

Terranova, the Actis-backed Latin American hyperscale platform, opened its Queretaro campus in Mexico in the first quarter of 2026 inside a $1.5 billion three-country rollout, with Campinas, Brazil scheduled for 2027 and Chile between 2027 and 2028. The platform was explicitly designed around sustained mid-tier inference rather than burst training, with clean-power siting and a 1 GW potential cap on the regional portfolio. The June 22 partnership announcement from BNDES, Vale and Petrobras on rare earth research, development and innovation, made formal at a meeting chaired by BNDES president Aloizio Mercadante, sits underneath this. The Brazilian materials supply chain now has a place inside a global compute build where open weights and cheap inference are the default rather than the exception.

US angle

The S-1 vintages look early, not late. Both OpenAI and Anthropic filed before the price-per-token chart visibly broke against them. A D.A. Davidson analyst quoted in the June 26 CNBC piece warned the largest enterprise customers may begin capping token spend outright. Microsoft’s MAI-Code-1-Flash launch June 2 was framed inside the keynote as a cost-reduction move that lowers reliance on OpenAI inside Copilot. The June 12 export-control directive that disabled Anthropic’s Fable 5 and Mythos 5 models for all foreign nationals, including foreign-national Anthropic employees, removed a top-tier closed model from non-US enterprise procurement on roughly thirty days notice. Anthropic’s public statement called it a misunderstanding and said it was working to restore access. As of late June, the restriction remained in place on Fable 5.

China angle

The GLM 5.2 release on June 13 is the first open weight model to clear a frontier closed model on FrontierSWE within a single point, at roughly a fifth of the inference cost. The Intelligence Index v4.1 ranking puts Zhipu (51) clearly ahead of MiniMax M3 (44), DeepSeek V4 Pro (44) and Kimi K2.6 (43). The June 22 China Ministry of Commerce action that placed MP Materials, USA Rare Earth and eight other US firms on the export control list landed inside the same week. Read together, the sequencing looks deliberate. Block US rare earth metallization on the materials side, push open weights at a fifth of US inference cost on the model side, force US frontier developers to defend a near-trillion-dollar valuation while Chinese open weight competitors price the floor.

What it means

The token price collapse is not a problem for the data center build. It is the demand for it. Cost per token fell roughly fourfold in a year, but enterprise AI usage is growing faster than that, so total inference volume is rising on net. This is the Jevons paradox version of inference economics. Cheaper tokens mean more tokens, which mean more hours on racks, which mean more racks, which mean more copper, gallium and uranium per dollar of GDP that runs through a hyperscaler. The materials read does not weaken after Friday. It reshuffles. The compute build that wins is the one positioned for sustained mid-tier inference at high utilization, not the one positioned for burst training peaks. Brazil’s first-phase data center geography fits that profile.

The valuation read is separate. OpenAI and Anthropic priced their S-1s before the open weight ceiling moved. Closed-frontier subscription margin is what supports a near-trillion-dollar print. The June 26 reporting suggests that margin is now under enterprise renegotiation. The S-1 vintage is the risk, not the build.

What to watch

Next Artificial Analysis Intelligence Index update. If the gap between Opus 4.8 and GLM 5.2 closes to zero on FrontierSWE, the closed-frontier subscription thesis breaks first there.
OpenAI and Anthropic S-1 amendments. Token economics disclosure is the new diligence line. Expect the first quantitative reveal in the first amendment cycle between July and September.
The next US export control update on the Fable 5 and Mythos 5 list. If the directive expands to lower-tier Anthropic or OpenAI closed models, the open weight floor lifts further and the GLM 5.2 procurement window widens.

Tantalum Strategy