AI Compute Supply 2026: Power vs Baseline Token Costs
What is the AI compute supply scenario?
The compute supply scenario is a new module on the AI Agency Pricing Calculator that prices supply-side risk the calculator previously ignored: it compares your baseline per-token cost against a power-constrained outlook for 2026–2027, with grid/power pass-throughs as an ESTIMATE add-on and locked-in capacity (up to 5GW AWS Trainium, ~3.5GW Google TPU from 2027) as an ESTIMATE discount. The driver is power, not chips: Gartner forecasts global data center electricity up 26% in 2026 to 565 TWh and past 1,200 TWh by 2030, where grid supply may be insufficient. The output is the monthly and 12-month delta on your token bill if that constraint binds — the number to carry into fixed-fee client quotes.
Token pricing is downstream of compute and power economics — and as of August 2026, both are being locked in years ahead. On August 10, 2026, Anthropic, Macquarie Asset Management, and Singapore's GIC announced Theseus Infrastructure, a platform to develop, operate, and lease purpose-built U.S. data centers to Anthropic under long-term agreements, with Anthropic as anchor tenant and Macquarie/GIC owning the platform and funding the majority of each project's equity. That same week, Anthropic pledged to pay 100% of the grid-upgrade costs of interconnecting its data centers. Meanwhile Anthropic has committed more than $100 billion over ten years to AWS for up to 5GW of Trainium capacity, and a separate Broadcom/Google deal adds ~3.5GW of TPU capacity from 2027.
For AI agencies pricing client spend, the story is not that Anthropic is leaving AWS (it isn't — AWS remains its primary training cloud). The story is that AI supply is being locked in years ahead, and power — not chips — has become the binding constraint. The calculator prices today's inputs only. Nothing on the page modeled supply-side risk: power caps, dedicated-capacity deals, or grid-upgrade pass-throughs. The compute supply scenario closes that gap.
Why power is now the binding constraint
The data point to anchor on is Gartner's forecast — and it is a forecast, not a measured figure. Gartner expects global data center electricity consumption up 26% in 2026 to 565 TWh, with capacity rising from 104 GW in 2025 to 132 GW this year, and total consumption passing 1,200 TWh by 2030, where "grid supply may be insufficient." Gartner's own framing: "AI capacity is now constrained by power availability, making datacenter power security the new battle ground for scaling and protecting margins."
That framing is why Anthropic's 100% grid-upgrade pledge matters. To be precise: the pledge is an Anthropic commitment, not a Theseus one. Anthropic says it "will pay for 100% of the grid upgrades needed to interconnect our data centers, paid through increases to our monthly electricity charges — including the shares of these costs that would otherwise be passed onto consumers." For agencies, the mechanism is the lesson: power costs are already being structured as pass-throughs. When a lab absorbs grid-upgrade cost through its electricity charges, that line eventually shows up in the cost of serving tokens.
The capacity already locked in
| Commitment | Capacity | Timing |
|---|---|---|
| AWS Trainium (Anthropic >$100B / 10yr) | Up to 5 GW | Ramping; Amazon $5B now + up to $20B on top of $8B in |
| Google TPU via Broadcom (Anthropic) | ~3.5 GW | From 2027 (on top of 1 GW in 2026) |
| Theseus Infrastructure (Macquarie/GIC owned, Anthropic anchor tenant) | Not disclosed | Long-term leases; no capital/capacity/lease terms published |
What the committed numbers mean: a meaningful share of frontier training and inference capacity for 2026–2027 is already sold to one anchor customer. Dedicated-capacity deals like Theseus and the hyperscaler commitments reduce the scarcity premium on that locked capacity — the supply-side reason token prices could hold or fall even as demand explodes. But they do not reduce the power cost underneath, and they do not disclose what Anthropic actually pays per watt or per token.
What this does to per-token costs in 2026–2027
There are two forces pushing per-token cost in opposite directions, and the scenario models both as user-set estimates:
- Grid / power pass-through (% add-on). The share of power-cost and grid-upgrade increases passed into per-token pricing. Direction: up. The 2026 preset uses 8%; the 2027 preset uses 12%, reflecting a tighter grid as Gartner's curve approaches 1,200 TWh.
- Locked-in capacity relief (% discount). The scarcity premium already removed by committed capacity — up to 5GW Trainium plus ~3.5GW TPU from 2027. Direction: down. The 2026 preset uses 0% (capacity still ramping); the 2027 preset uses 5% (TPU capacity online).
Run the math on a representative agency: 100 million tokens/month at a $3.00 blended baseline. Baseline monthly cost: $300. Power-constrained 2026 (8% pass-through, 0% relief): $324/month — a +8% move, +$288 over 12 months. Power-constrained 2027 (12% pass-through, 5% relief): $319/month — a +6.4% move, +$230 over 12 months. Small percentages, but they compound across every client contract you quote on token-based pricing.
These percentages are estimates, not published prices. No capital commitment, capacity, site count, or lease terms were disclosed for Theseus by any party. Amodei's buildout math ("hundreds of billions, not trillions"; roughly $10–15 billion per gigawatt on a 10–15 GW industry buildout this year) is his framing on the Dwarkesh podcast, not a published cost schedule. The scenario exists so you can substitute your own assumptions and stress-test the direction, not so it can hand you a precise future price.
Run the scenario, then export the PDF
Open the AI Agency Pricing Calculator, scroll to the Compute Supply Scenario section, enter your monthly token volume and blended baseline cost per 1M, pick a horizon, and adjust the two ESTIMATE sliders to your own read of the power market. The results panel shows baseline vs power-constrained monthly cost, the delta, the effective per-1M rate, and a 12-month outlook — and the Print / Save PDF button exports the whole panel for a client quote or your own pricing review.
Model your AI delivery costs against current market rates — then stress-test supply
Run the Compute Supply Scenario →Estimate setup fees, retainers, and margin in under a minute — then compare per-agent workload costs with the AI Agent API Cost Calculator.
Frequently asked questions
What is the AI compute supply scenario?
It is a new scenario module on the AI Agency Pricing Calculator that compares your baseline per-token cost against a power-constrained outlook for 2026-2027. It models supply-side risk the calculator previously ignored: grid/power pass-throughs (an ESTIMATE add-on) and locked-in capacity relief (an ESTIMATE discount from already-committed Trainium and TPU capacity). The output is the monthly and 12-month delta on your token bill if power constraints bind.
Why is power the binding constraint on AI compute in 2026?
Gartner forecasts global data center electricity consumption up 26% in 2026 to 565 TWh (capacity rising 104 GW to 132 GW), passing 1,200 TWh by 2030 where grid supply may be insufficient. Gartner's own framing: AI capacity is now constrained by power availability, and power access has become the real scarcity — which is why Anthropic pledged to pay 100% of grid-upgrade costs for its own data centers (an Anthropic commitment, not a Theseus one).
What capacity is already locked in for AI compute?
Anthropic has committed more than $100 billion over ten years to AWS technologies for up to 5 GW of Trainium capacity, and a separate deal with Google and Broadcom gives it approximately 3.5 GW of next-generation TPU capacity from 2027. AWS remains Anthropic's primary training and cloud provider. Theseus Infrastructure (Anthropic, Macquarie Asset Management, GIC — announced Aug 10, 2026) is dedicated capacity financed by institutional capital, but no capital commitment, capacity, or lease terms were disclosed.
How do I run the compute supply scenario?
Open the AI Agency Pricing Calculator, scroll to the Compute Supply Scenario section (or click the FAQ link "Run the Compute Supply Scenario above"), set your monthly token volume and baseline cost per 1M tokens, pick a horizon (2026 or 2027), and adjust the two ESTIMATE assumptions — grid/power pass-through % and locked-in capacity relief % — then click Run Compute Supply Scenario. Use Print / Save PDF to export the result.
Sources
- Macquarie Asset Management — Theseus Infrastructure press release (Aug 10, 2026): macquarie.com
- Anthropic — Covering electricity price increases from our data centers: anthropic.com
- Data Center Dynamics — GIC and Macquarie form Theseus Infrastructure (Aug 11, 2026): datacenterdynamics.com
- The Real Deal — Anthropic recruits Macquarie, GIC for data center venture (Aug 11, 2026; snippet corroboration only — full text blocked): therealdeal.com
- AIWeekly — Anthropic, Macquarie and GIC launch Theseus for AI data centers (Aug 10, 2026): aiweekly.co
- Xinhua — Anthropic, Macquarie, Singapore's GIC partner on U.S. AI data centers (Aug 11, 2026): english.news.cn
- Converge Digest — Anthropic Backs Theseus Infrastructure: convergedigest.com
- best-ai.org — Anthropic, Macquarie, and GIC Launch Theseus Infrastructure (Aug 11, 2026): best-ai.org
- Anthropic — Anthropic and Amazon: expanding our partnership (Apr 20, 2026): anthropic.com
- Amazon — Amazon invests additional $5 billion in Anthropic (Apr 20, 2026): aboutamazon.com
- TNW — Anthropic's Google/Broadcom compute deal, ~3.5GW TPU (Apr 7, 2026): thenextweb.com
- The Register — Datacenter growth may run into a power wall by 2030 (Jun 10, 2026): theregister.com
- Dwarkesh Podcast — Dario Amodei (buildout math: "hundreds of billions, not trillions"; ~$10–15B/GW): dwarkesh.com
Accuracy note: Gartner figures (565 TWh 2026, +26% YoY, >1,200 TWh by 2030, grid possibly insufficient) are forecasts and are labeled as such. The 100% grid-upgrade pledge is an Anthropic commitment (its own newsroom page), not a pledge by the Theseus venture entity — it appears in the Macquarie PR as covering consumer electricity-price increases, and is absent from the DCD and Real Deal coverage. No capital commitment, capacity, site count, or lease terms were disclosed by any party for Theseus. The $10–15B/GW buildout cost and "hundreds of billions, not trillions" figures are Dario Amodei's framing on the Dwarkesh podcast. The Real Deal article is 403-blocked for full-text access; facts attributed to it are from its indexed snippet only. The scenario's pass-through and relief percentages are estimates, not published prices — substitute your own assumptions before quoting client work.