Blogs · Briefing for CIOs, CTOs and CPOs · November 2026
Your AI bill is about to become a multi-year contract.
Frontier labs are locking in GPUs for years and will ask you to do the same with tokens. AtriumLLM helps you sign those commitments at the right size, use them on plan, keep overage priced as overage, and stay free to switch.
The signal
A VC reading Anthropic's S-1/A says the next AI winners will be decided by contract terms.
“…match the duration of their customers’ contracts to the duration of their compute.”
Tomasz Tunguz, Theory Ventures, “The Financial Engineering Essential for AI in 2027”, Nov 3, 2026
His argument in brief: when a lab's long-term GPU capacity runs out, it buys short-term capacity at several times the price. Raising token prices is risky in a price war, so labs are responding with efficiency gains, tighter subscription allowances, and a push to move enterprise buyers from metered API spend into multi-year minimum commitments at locked-in rates.
Short-term GPU rental cost in Anthropic's S-1/A, against $2–4/hr on long-term contracts.
Estimated blended cost per GPU-hour, March to September 2026.
Share of data center projects running late, per Goldman Sachs research cited in the post.
Cost per task for Opus 5.5 compared with Opus 5, as labs offset GPU costs with efficiency.
Share of Anthropic's Q3 2026 revenue that is consumption-based rather than subscription.
Figures as reported in Tunguz's post and its cited filings; the blended GPU-hour figure is his estimate within a stated range.
What changes for you
Four risks land on the buyer, and AtriumLLM is built to manage each one.
A multi-year commitment can lower your unit price. It also moves forecasting, utilization and switching risk from the lab's balance sheet to yours. AtriumLLM sits in the request path between your applications and every model provider, so it can act on that risk request by request.
Price swings you can't forecast
Allowances get pooled, rate cards change, and burst costs leak into list prices.
Every request is priced at the moment it runs and recorded with its model, provider and cache status.
- Cost per task, not just cost per token, so a cheaper token that needs more of them shows up as more expensive.
- Cheaper lanes for routine work: open-weight models and small fine-tuned models handle classification, extraction and drafting.
- Session-level routing that keeps prompt caches warm, often a bigger saving than switching models.
Committed spend you don't use
You pay the full minimum whether or not the tokens get used.
Commitments are first-class objects in the routing registry, and eligible traffic uses committed capacity first.
- A burn ledger that reconciles daily with provider invoices.
- A forecast for each commitment period with alerts at under-use and over-use thresholds.
- An opt-in burn-down queue that points deferrable work, such as evals and backfills, at an under-used commitment.
Overage at burst rates
Exceed the commitment and the extra tokens bill at on-demand or worse.
Overflow is a routing decision with a reason attached, not a surprise on the invoice.
- When a commitment nears exhaustion, traffic moves to other eligible commitments before on-demand capacity.
- Every overflow request records why it overflowed and what it cost compared with the committed rate.
- Regulated workloads fail closed rather than overflow to a non-compliant provider.
Lock-in at renewal
If you can't move volume, you can't negotiate the rate.
Applications call AtriumLLM, not a vendor SDK, so switching providers is a registry change rather than a code change.
- A portability score: the share of your traffic that already has a tested cross-vendor fallback.
- A renewal pack with 12-month usage, growth, and cost per task by provider, ready for procurement.
- One control plane across direct provider contracts, cloud marketplaces and model marketplaces.
One rule never bends: AtriumLLM chooses which contract pays for an approved model. It never swaps in a weaker model to hit a spend target.
What your CFO sees
Commitment burn against plan, with the gap visible months before quarter-end.
By role
What it means for each of you.
Size the commitment from evidence
Base the commit on your own traced usage and growth by business unit, then hold every unit to its share with chargeback your finance team can audit.
Keep architecture portable
Applications call one gateway. Model choice, fallbacks and provider contracts live in a reviewed registry, so a renewal never forces a rewrite.
Protect product margins
Know the cost per task of each AI feature, set quality floors per feature, and price your own plans knowing where your upstream costs are headed.
Before you sign
Five questions to answer before any multi-year token commitment.
- What is our real baseline, and what is burst?Commit to the steady load. Leave launches, experiments and seasonal peaks on flexible capacity.
- Which workloads must stay on this vendor, and which could move?The share that could move is your negotiating leverage. Measure it before the first call.
- What happens to unused volume?Ask about rollover, true-up timing, and whether cache reads and reasoning tokens count toward the minimum.
- How is overage priced, and who sees it first?An overage clause is only useful if someone is alerted before it applies.
- Can we exit if the market moves?Supply may ease in 2028. A rate locked high needs a renegotiation point or a step-down clause.
Get a commitment readiness review before your next AI contract.
We'll map your current model spend by workload, estimate a defensible committed baseline, and show how much of your traffic could move between providers today. Join early access and we’ll reach out.
Join early access →- Tomasz Tunguz, “The Financial Engineering Essential for AI in 2027,” Theory Ventures, Nov 3, 2026, and the filings and research it cites (Anthropic Form S-1/A, Goldman Sachs Commodities Research).