The facts: Three moves in seven days are redrawing the AI value chain. On 22 June, SpaceX acquired Cursor, the world's most widely used AI development tool, for 60 billion dollars, and announced a 75-billion IPO. On 24 June, OpenAI and Broadcom unveiled Jalapeño, OpenAI's first custom chip, with a commitment to deploy 10 gigawatts of capacity at Microsoft and other partners by 2029. At the same moment, Anthropic entered advanced talks with Microsoft to run Claude's inference on the Maia 200 chip — dedicated silicon, launched in January on TSMC's 3 nm process, promising more than 30% better performance per dollar over general-purpose chips. Meanwhile, Anthropic is accusing Alibaba of orchestrating a massive distillation operation (25,000 fraudulent accounts, 28.8 million Claude exchanges between April and June).

The Codito analysis: For three years, the model layer soaked up all the attention: GPT-5, Claude, Gemini. The layer beneath — the chips — remained the blind spot. This week, it became the main battleground, and every AI provider is now looking to control the entire chain: chip → model → tool. OpenAI is bringing its silicon in-house. SpaceX is absorbing the dev tool that reaches developers. Anthropic is negotiating inference on third-party silicon. Microsoft is running its own chips inside Azure and potentially hosting Claude on them. This is no side story: it marks the end of a market where you could naively pick a model and forget what sat underneath. Tomorrow, your AI choice will implicitly be a full-chain choice — including the energy chain that powers it.

What it means for you: Three consequences for a French SME. On costs: dedicated inference chips (Maia 200, Jalapeño, Google's TPU) lower the marginal cost of a model call. That drop will cascade up to your enterprise subscriptions, most likely through price cuts or bigger included quotas, over 12 to 18 months. On reliability: less variance, fewer outages at peak hours. The gap between “what we can test” and “what we can deploy in critical production” is going to shrink. On sovereignty: compliance is no longer contracted at the model level alone. It plays out across the entire stack. Action to take this very week: add one simple question to your internal framework — what infrastructure will the agent we deploy actually run on?

The number of the week 10 gigawatts. That's the compute capacity OpenAI, Broadcom and Microsoft plan to deploy on Jalapeño chips by 2029. To make that tangible: it's the equivalent of the city of Paris's electricity consumption at winter peak. One single partnership. Just to run AI. AI is no longer a software topic — it has become an energy infrastructure issue.

At first glance, nothing changes for you: you use ChatGPT, Claude or Gemini, and their interfaces never mention chips. And yet: the most structural decision of the week — for the price, quality and dependency of your AI over the next 18 months — was made on the silicon side. Here's how to get ahead of it.

1. Costs are coming down — significantly

Chips dedicated to inference (Maia 200, Jalapeño, TPU) are designed to do one thing better than general-purpose chips: run AI models in production. “Better” means two to three times more compute per dollar spent. Those savings will travel upstream — Microsoft and OpenAI will be pushed by competitive pressure to pass them on. Expect price cuts of 30 to 50% on the enterprise versions of the models by late 2026 and early 2027. Plan your annual contracts accordingly — no blind automatic renewals.

2. Reliability is going up — and with it, critical use

Dedicated chips mean less variance in response times, fewer outages, more capacity available at peak hours. That's exactly what separates a “toy” AI deployment from a “critical” one. You'll soon be able to hand an AI agent a task you expect executed within the minute, without fearing it will crawl because a million people are querying ChatGPT at the same time. The gap between “what we can test” and “what we can deploy in production” is going to shrink. Organisations that have already scoped their priority use cases will be ready to switch from testing to critical without friction.

3. Your data sovereignty is changing layers

For three years, the reasoning went: “is my data in Claude Enterprise or ChatGPT Enterprise?”. The relevant question tomorrow: which chip does it run on. If Anthropic ends up running Claude on Maia 200, your data passes through Microsoft's chips, even if you're an Anthropic customer. If OpenAI deploys Jalapeño with Broadcom, that's yet another link in the chain. Sovereignty is no longer contracted at the model level alone. It is chosen at the level of the full stack — chip included. This is now a live topic for organisations with heavy GDPR stakes (notaries, funds, healthcare, defence).

“For three years, you chose your model. Now, you choose your chain.”

The action to take this very week: add one simple question to your AI usage charter — what infrastructure will the agent we deploy actually run on? If your provider can't answer clearly, it's not the right one. If the answer is “we don't know, the partner changes every six months”, you're buying risk without knowing it.

Book a strategy call

Discover the Codito AI Partnership

To close this edition, an upcoming event — one that directly extends Parts 1 and 2.

On Tuesday 2 July, Émile, my partner and President of Codito Ergo Sum, co-hosts — alongside Julien Van Quackebeke (UpSkill4IT) — a talk at the 4^th edition of the Azur Tech Summer, organised by Telecom Valley at SKEMA Business School (Sophia Antipolis).

The exact title

“The Prompt → Context → Agentic evolution: what each wave brought, what it left unsolved, and where to set the cursor today.”

What Émile will bring to it

On the agenda: the hidden costs of each AI wave (model dependency, exploding API bills, black-box effects) and four durable disciplines for producing quality AI code — in short, exactly the kind of method we champion every Sunday in this newsletter. The angle: what these waves, seen from above, mean in practice for a French SME. How to position yourself today without committing to a stack that will be obsolete tomorrow — precisely the subject of Parts 1 and 2 of this edition.

It's the first time Codito takes a stage of regional significance. A chance to show that the AI Partnership method and the AI maturity grid — shared here every week — hold up outside a client call, and apply to every sector Telecom Valley represents.

If you're in the South of France — or will be in early July Come say hello. It's a real opportunity to talk face to face about your own AI trajectories, in an operational setting. Details and registration on the Telecom Valley website. We'll be there on Tuesday.