Skip to main content
biznesam.ai
AI Briefing

AI infrastructure splits from AI usage — three signals in one day

Microsoft starts counting tokens, OpenRouter raises $113M for routing, SoftBank pours €75B into French data centers. All three stories are one.

LV EN

Summary

  • Microsoft starts counting tokens — GitHub Copilot moves to usage-based billing on June 1. Flat $10/month is gone.
  • OpenRouter raises $113M — $1.3B valuation for a platform routing 400 AI models. Google and NVIDIA both invested.
  • SoftBank pours €75B into France — 5 GW of AI data center capacity in Hauts-de-France. Europe’s sovereign AI signal.

Bottom line: AI infrastructure and AI usage split today. The model becomes a commodity; the router and sovereign capacity become the new moat.

1. Microsoft starts counting tokens — the end of Copilot flat pricing

What changed. Starting June 1, 2026, all GitHub Copilot plans move to usage-based billing — $10/month includes a fixed allotment of AI Credits; beyond that, you pay per token at API rates (TechCrunch). The official GitHub community thread collected 900+ downvotes and 400+ comments. One developer estimated that a single agent session burns $30-40 — a Pro user with $10/month in credits hits the ceiling in one working day (GitHub Blog).

Why it matters. This is the first major signal that the “AI tools as gym membership” era is over. Microsoft is exposing the real economics — an agent session isn’t free, and it costs more than you thought. Every other AI vendor will see this and follow. Your finance team currently has no idea what AI will cost next month.

What to do this month.

  • Ask your IT lead: how many tokens per month does our dev team consume across Copilot/Cursor/Claude? If they don’t know — restrict access until you have measurement.
  • Audit every AI tool subscription for usage-based pricing risk — not just Copilot. Notion AI, Linear AI, Slack AI are next candidates.
  • Add a budget cap at the API key level for every AI tool. A simple usage cap insures against billing surprises.

What I expect.

  • By August, expect 3-5 more major AI tools to announce a switch to usage-based pricing (likely candidates: Cursor, Codeium, Vercel v0).
  • By year-end, an “AI cost observability” tool market emerges — startups tracking token consumption at enterprise level (Helicone, Vellum already growing).
  • In the next 90 days, a major company will publicly disclose unexpected 5-6 figure AI bills. That will shift CFO attitudes toward AI budgeting.

2. OpenRouter raises $113M — routing is the new business

What changed. OpenRouter, a unified API platform for 400+ AI models from 60+ providers, announced a $113M Series B at a $1.3B valuation (TechCrunch). Investors: CapitalG (Google), NVentures (NVIDIA), Andreessen Horowitz, plus Snowflake, MongoDB, Databricks corporate funds. Weekly throughput grew 5× in six months to 25 trillion tokens — clients include Replit, Framer, Zoom, Webflow.

Why it matters. Google and NVIDIA — two competitors in the models/hardware market — are jointly investing in a platform that routes requests between their models and their competitors’. That’s a powerful economic signal: models are becoming commodity; the profit pool moves to orchestration. If your team is tightly coupled to one AI vendor (OpenAI or Anthropic), this month you’re seeing that architecture lose value.

What to do this month.

  • Audit: how many AI vendors have access to your business-critical processes? If one — you have vendor lock-in risk.
  • Ask your tech lead: does our code use a single model SDK (e.g., openai), or an abstract routing layer (OpenRouter, LiteLLM)? In the first case, switching costs grow over time.
  • If you don’t yet use multi-vendor routing — pick one experiment (e.g., customer support or document processing) and switch to OpenRouter this month to measure cost differences.

What I expect.

  • In the next 60 days, several companies will declare “multi-model architecture” as a 2026 principle (similar to “multi-cloud” in 2018).
  • By Q4, OpenRouter competitors (LiteLLM, Portkey, Anthropic Workbench) will announce larger rounds — the category is growing.
  • Average per-token model prices will keep falling 30-50% annually, but total enterprise AI bills will rise (Jevons paradox — cheaper means more consumption).

3. SoftBank pours €75B into French data centers — Europe’s sovereign AI

What changed. SoftBank announced a €75B investment in French AI data centers — 5 GW of total capacity by 2031 (Fortune). First phase: €45B and 3.1 GW in the Hauts-de-France region (Dunkirk, Bosquel, Bouchain). Announced at Macron’s “Choose France” summit. Partners: EDF (electricity) and Schneider Electric (factory automation).

Why it matters. This is the largest European AI infrastructure commitment from a single firm. Macron is positioning France as the alternative to the US and China in the sovereign AI market. For Baltic and European companies this means: real EU-based AI cloud providers will arrive in the next 2-3 years — usable without transferring data to the US, a critical requirement for GDPR and NIS2 compliance.

What to do this month.

  • If your AI usage faces data residency requirements (health, finance, government clients) — demand a roadmap from your vendor showing EU data centers (not US/UK).
  • Audit current AI contracts — where exactly is data processed? Many “European” AI solutions today call US APIs under the hood.
  • Keep a 12-month window for choice: commitments made now with US-based AI vendors could become a migration burden when competitive EU alternatives arrive.

What I expect.

  • By mid-2027, at least 2 real, functioning EU sovereign AI offerings will be available (Mistral as model + French/German hosting).
  • In the next 6-12 months, large EU enterprises (banking, telco) will start requiring “EU sovereign AI” clauses in contracts.
  • For European companies with EU client portfolios, this becomes a sales argument — “we process your data only on EU infrastructure”.

Today’s picture

All three stories are one story: the AI market split into three layers today. Bottom — physical infrastructure (SoftBank, data centers, sovereignty). Middle — models becoming a commodity bundle (OpenRouter routing). Top — applications whose real economics are now visible (Microsoft counting tokens). The critical question for any leader: which layer do you compete in? If models — you won’t earn margin. If applications or data — you need to know your token cost structure this month.

EventConsequence
Microsoft switches to usage-based pricingAI costs become unpredictable; a budget control layer is required
OpenRouter $1.3B valuationModels are a commodity; multi-vendor architecture becomes standard
SoftBank €75B in FranceEuropean sovereignty infrastructure real by 2027-2028; plan the transition

3 questions for leadership:

  • How many tokens per month does our company consume now, and what is its growth rate?
  • Is our AI architecture locked to one vendor — and what happens if they double the price?
  • Will our clients soon demand EU data residency for AI processes — and can we deliver it?

Sources