Three Revealed-Preference Events in One Week
Microsoft cancelled internal Claude Code licences this week.
Token-based billing made the cost untenable — even for the company that committed about $13 billion to OpenAI and runs the Azure capacity powering most of Anthropic's compute. If the bill is too high for Microsoft, it is too high for everyone below Microsoft on the demand curve. This is a price-discovery event for the whole enterprise market, not a vendor-relationship event.
Uber's CTO sent an internal memo: entire 2026 AI budget consumed in four months.
Four months means the annualised run-rate is roughly three times the planned spend. The budget had been set on the assumption that effective AI costs would keep falling, not rise. This is the first publicly-leaked CFO-side confirmation that enterprise budgeting models built on the 2023–2025 cost-decay assumption no longer hold.
American AI software prices +20–37%; GitHub drops flat-rate plans.
GitHub — owned by Microsoft — is phasing out flat-rate plans in favour of usage-based metering across Copilot, Codespaces, and Actions. Microsoft is migrating its own customers onto usage-based billing while simultaneously refusing to absorb usage-based billing from a competitor. The pricing-model convergence on metering is now industry-wide.
The Take
Anthropic, OpenAI, and Google all raised effective prices in the last six months. Not by raising sticker rates — by raising per-token consumption through longer context windows, reasoning-mode token expansion, and tool-use overhead. The sticker-price stability hid the per-task cost growth. Uber's four-month budget overrun is the first visible point on that curve crossing the enterprise budgeting threshold.
Two outcomes look likely from here. Either enterprises scale back AI usage to fit budgets, which slows the revenue ramp the labs need to justify their valuations ahead of expected 2026–2027 IPOs. Or the labs cut prices and absorb the losses, which makes the unit economics worse at exactly the wrong moment. Both paths land in the same place: the numbers stop working, and somebody takes the writedown.
Token-based pricing is forcing every enterprise customer to confront the actual cost of running these models at scale. The number turns out to be far higher than the flat-rate experiments suggested.
Why Our Approach Is Solid
DWS IQ 6 was designed under a different hypothesis: that compute would become scarce, sovereignty would become bankable, and the per-token meter would become the operating-cost surface enterprise CFOs could no longer accept. We built the platform on owned EU-sovereign compute — not because we predicted week 21 of 2026, but because the architecture made sense on its own terms. The market just walked into it.
The NVIDIA-Models Moat Beneath the Envelope
The reason we can quote a flat annual envelope is that the inference layer underneath it does not run on someone else's token meter. It runs on NVIDIA hardware that we own, with model weights that the open-source community produced at NVFP4-native precision at 500B scale.
| Layer | What runs there | Why it is a moat |
|---|---|---|
| Hardware | Owned NVIDIA DGX Spark, EU-located, deterministic monthly cost | Hardware tick, not token tick — cost cannot drift on us |
| Brain (Primary) | Nemotron 3 Super — 123.6B MoE / 12.7B active, Mamba-2 + attention + Latent MoE, NVFP4-pretrained at 500B scale | ~4.5× decode speedup vs Llama 70B baseline; native NVFP4 means no quantisation tax |
| Brain (Fast) | Qwen3-Coder 30B on the same Spark fabric | Code/tool tasks served locally; no remote billing on tool-use overhead |
| Runtime | OpenClaw on owned compute — agent shell, file I/O, mailbox, browser, all in MicroVM | Tool calls do not invoke external APIs; per-action cost is zero at the meter |
| Memory | Prompt caching architecture against the owned cache | Long-context (regulatory documents, audit logs) paid once, refreshed locally |
| Fallback | Multi-provider routing — Groq, Baseten, Anthropic, OpenAI — task-aware model-selector | If a frontier price moves the wrong way, we route around it; we are never locked to one curve |
This is the §14.18 supply-side companion to the §14.20 demand-side signal: compute scarcity made owned EU-sovereign Spark the right hardware bet; the NVFP4 pretraining at 500B scale made Nemotron 3 Super the right model bet on that hardware; the demand-side capitulation made the resulting fixed-cost envelope the right commercial bet. Three independent moats stacked. Competitors running on Anthropic or OpenAI passthrough do not have the hardware layer, do not have the NVFP4-native model layer, and therefore cannot quote an envelope that holds when the upstream meter moves.
We did not switch off the frontier labs. We routed past them. The frontier-lab API is one tool in our hands; the runtime is ours.
The asymmetry is structural, not tactical:
| Demand-side pressure | DWS IQ 6 (owned compute) | Per-token competitor |
|---|---|---|
| Inference cost | Fixed amortised hardware tick (Spark NVFP4, Nemotron 3 Super, Qwen3-Coder) | Token meter, vendor-controlled growth |
| Long-context workloads | Prompt cache on owned hardware — repeated context paid once | Re-billed every call |
| Tool-use overhead | OpenClaw runtime on owned hardware — no remote billing | Each tool call invokes the meter |
| Agent identity (mailboxes, channels) | Proton mailbox-per-agent, flat-rate plan | Per-message AI cost |
| Effective price drift | Multi-provider routing — Groq / Baseten / Anthropic / OpenAI alternated by economics | Locked to one vendor's pricing curve |
| Adding more agents | Amortised hardware-tick growth | Multiplied token consumption |
For procurement conversations, this means we can quote an annual envelope anchored on hardware-amortised inference cost — not a per-task quote anchored on speculative token consumption. Competitors running on Anthropic or OpenAI passthrough cannot replicate this without taking the per-token gap directly onto their own income statement.
The Compliance Angle: Cost-Surface Stability
The EU AI Act Article 14 requires human oversight to remain effective throughout deployment. But if a deployer's AI budget collapses mid-year and oversight is cut to fit the residual envelope, the system is no longer running in the configuration on which conformity was assessed. The cost-surface of an AI deployment is now a conformity-stability variable that Article 14 does not currently capture.
We are adding this as a third conformance track to the KYA Specification candidate alongside the existing two: ASB-track (operational safety, OWASP Top 10 for Agentic Applications) and EAB-track (decision quality, EAC1–EAC13). The new CSB-track (Cost-Surface Stability) extends Article 14 oversight from a quality variable to an affordability-stability variable. It is productionisable on owned-compute inference. It is not productionisable on metered-API inference.
ASB governs what agents do. EAB governs how agents choose. CSB governs what running the agents costs the deployer over the conformity-assessed period.
What Buyers Should Ask
Can you quote me an annual envelope I can take to my CFO?
Owned-compute vendors answer with a number. Per-token vendors answer with a forecast that depends on vendor-side pricing changes they don't control.
If your upstream provider raises per-token prices 30% in six months, what happens to my contract?
Ask in writing. If the answer is "we pass it through," you are absorbing the next round of the pricing capitulation. If the answer is "we run on owned compute," you are not.
If my budget collapses mid-year, does your platform still meet Article 14 oversight?
Cost-surface stability is the silent assumption inside EU AI Act conformity. Cut the budget, cut the oversight, lose conformity. CSB-track makes this auditable; owned-compute makes it real.
Match the Solution to a Product
Three configurations, all priced as a one-time deployment, all running on owned NVIDIA Spark inference with NVFP4-native Nemotron 3 Super and Qwen3-Coder 30B underneath. No per-token meter on the inference, the mailbox, the tool-use, or the prompt cache:
Platform · DWS-IQ-006
DWS IQ Platform v6
Price list · one-time
A fleet of skilled agents with agentic loops across 25 EU-regulated industries. MCP-native, CLI-accessible, EU-hosted on owned Spark compute. Cuts 90% of manual compliance hours.
Right for: teams that want the whole platform on a flat envelope.
Bundle · LIFETIME-BUNDLE-2026
Climate Reporting Starter Bundle
Price list · one-time
DWS IQ Base Platform + Firehorse Report Module + Kickoff Workshop. Built for the first CSRD reporting deadline. Operational in 30 days on owned EU-sovereign compute.
Right for: first-time CSRD filers who want zero-to-compliant in a month.
Automation · FH-AUTO-001
Firehorse Amplify
Price list · per project
Event-driven compliance workflows for Fit for 55, CSRD, and EPBD 2028. Captures ESG data directly from source systems with a complete audit trail — on the same owned-compute fabric.
Right for: teams that want the data pipeline locked before scaling the platform.
Annual envelope on the price list · Owned NVIDIA Spark + NVFP4-native Nemotron 3 Super · No per-token meter
Author: Risto Anton Päärni, Founder, CEO, Lifetime Oy
Published: 21 May 2026, Helsinki
Scientific framing: PHD_FIREHORSE_EN.md §14.20 (AI Pricing Capitulation), Key Contribution #25 (Demand-Side Capitulation and the Cost-Surface Stability Dimension of Agentic AI Conformity).
Read Also
Market Signal
The Moat Is Not the Model — NVIDIA Makes It Possible, EU-Sovereign Deployment Makes It Defensible
Supply-side companion to this piece. Compute scarcity and the NVFP4 hardware moat behind owned EU-sovereign inference.
Field Note
Baseten Raises $300M — Inference Engineering Is a Discipline Now
Why the inference layer is now a category. The cost structure beneath every token-pricing decision.
Buyer's Note
What You Should Know About Harness Engineering
The earlier buyer's-note framing. Read together with this piece to understand both the capability and the cost-durability dimensions of the procurement decision.