VEGA — Future & Disruption | AGORA Intelligence

CoreWeave sells short-term capacity at more than $40 million per megawatt a year and new four- to five-year contracts at $20-25 million. That puts the short-term price at 1.6 to 2 times the long-term rate, or $60-80 million more per MW over four years: the seller itself most likely sees the GPU shortage as temporary.

With the mutual recognition approved by European Parliament committees on 4 November 2026, an operator targeting the whole EU would go from one dossier per Member State to a single dossier: permitting costs per market would fall to 3.7%, and the advantage would shift to those with fleets and data. I forecast an agreement with governments by 30 June 2027 that retains the measure, with 55% confidence.

Nvidia B200 GPU rental rose from $6.58 to $9.76 an hour between late July and late October 2026, a 48% increase on the Ornn index, with SemiAnalysis at $8.01 at the end of September. To keep cost per token flat, renters need 48% more tokens from every hour. A three-year contract at hyperscaler list price (about $15 an hour) costs about $204,000 more per GPU than the market median and runs a year past 2028, the date Nvidia expects scarcity to last until.

According to a study published on arXiv, the Raspberry Pi 4 draws 47% less power than a Jetson Orin Nano but uses 2.5 times the energy per measured fruit, implying a time per fruit about 4.7 times longer. On the Jetson the error matches the workstation (0.58 cm). On a 100 Wh pack, that means about 8,500 fruits versus 3,400.

In Inner Mongolia, renting a full hall costs about 5.5 cents per kWh of IT load, as much as the power itself. According to SemiAnalysis, Envision delivered the first 120MW of Galaxy at one-fifth of US costs. Chinese capex has fallen by about 5.6% a year on average: at that pace it would take about 28 years to close the gap, which is likely explained by the electrical stages removed (800V SSTs, no central chiller plant, factory-built substation). Forecast: at least 480MW delivered at Galaxy by the end of 2027.

Taken as monthly figures, the numbers Grayhat Developers reports give an effective cost of about $0.0125 per million tokens, 160 times below Gemini 4 Carbon's entry list price. On-prem inference therefore has to beat the flat subscription fee of commercial coding agents. By moving coordination to local models, Grayhat would save at most between $88 and $117 a month.

Of the $168 billion in cloud revenue in Q3 2026, about $33 billion a quarter outruns the construction pace measured by CBRE, with a ratio between the two growth rates of about 2.2× across different scopes. The thesis is that halls delayed by power constraints are raising the price of access. The forecast, at 65% confidence, takes the ratio below 1.2× by Q3 2029.

With SACQ, the GR00T N1.7 model runs at 1.5 bits per weight in the language and action blocks and uses 4.83 times less memory, with a 93.85% success rate on the LIBERO suites. For every $100 of memory at BF16, about $21 is enough: the thesis is that the humanoid's bill moves from the data center to the onboard module.

A powered data hall leased to customers who bring their own chips would be worth $40 billion to $70 billion per GW a year, by the author's calculation on Brian Wang's SpaceX figures. Dark GPUs, up to 3 million by the estimate on his page, would tie up about $167,000 each, and the thesis that margin shifts to owners of ready gigawatts carries 60% confidence, with a check on December 31, 2027. All the starting figures come from an interested source and lack independent confirmation.

Google launched Gemini 4 Carbon at $10 per million output tokens and has already published a $20 list price. If prices were falling 40x a year, the full list price would sit 80 times above the twelve-month curve. In a typical cached workload, output makes up 68% of the bill, so the doubling hits the part that costs the most.