Infrastructure

The Nvidia Tax Just Found a Second Source

Meta's MTIA chip Iris enters production in September after clearing testing in six weeks. Why a hyperscaler's second source caps the price you pay for compute.

InfrastructureBy Servola Tech Desk2026-07-103 min read

AI-assisted, edited by humans. | Editor-in-Chief: Leon Soliman | Editorial standards

The Nvidia Tax Just Found a Second Source

A chip that floundered for years just cleared test in six weeks

An internal Meta memo reviewed by Reuters says the company's own AI accelerator, code-named Iris, starts production in September, and one striking line explains why this matters: at least one chip passed its testing phase in about six weeks with no major issues. Meta has been trying to build its own silicon since 2023, and the effort mostly stalled. A clean six-week bring-up is the kind of result that turns a research project into a supply line.

The part is a Meta Training and Inference Accelerator, the in-house family Meta designs to run the AI behind Facebook and Instagram. Broadcom helped design it, TSMC will manufacture it, Samsung supplies the memory, and the plan is to push Meta's compute from about seven gigawatts this year toward roughly double that next year. The chip itself is not the story. The speed of the test is.

The tell is that Meta keeps buying Nvidia anyway

Meta is not walking away from Nvidia. The same company putting its own chip into production still expects 125 to 145 billion dollars of capital spending this year, much of it on Nvidia and AMD hardware, and it has said plainly it will keep buying from both. That is the signal worth reading. A buyer this size does not build a second source to quit its supplier. It builds one to negotiate.

Every large component buyer knows the move. The moment a credible in-house alternative exists, the incumbent's pricing power changes, even for the units you still buy from them. Meta is not announcing independence from Nvidia. It is announcing that it now has a floor under its own costs and a card to play at the next price review, which is a very different and more durable thing.

What a second source does to the price you pay

You do not run a hyperscale fab, so the direct effect on your business is not the chip, it is the price of renting compute. Today that price carries a large premium because a single vendor sets it. When the biggest buyers start self-supplying part of their demand, that premium stops being a one-way bet, because the merchant market has to compete with the buyers' own silicon on cost.

The catch is timing. Meta's chip runs Meta's own workloads first, and custom accelerators are notoriously hard to rent as general-purpose capacity, so nothing on your invoice changes this quarter. What changes is the direction of travel: the assumption that GPU rental prices only rise, which underwrites a lot of 2027 planning, is now the assumption most exposed to being wrong.

What an operator should do before 2027 budgets close

Treat this as a budgeting input, not a hardware story. If your model for next year bakes in today's compute prices climbing at last year's rate, add a second scenario where the largest buyers self-supply and that curve flattens, and see which decisions flip. The point is not to predict the price. It is to stop betting the plan on one direction.

Then ask your cloud or GPU vendor a direct question at renewal: what happens to my rate if a major customer moves a chunk of its demand in-house. A vendor that shrugs is telling you the premium is still theirs to keep. A vendor that engages is telling you the second source is already priced in. Either answer is worth more than the headline about Meta.

Frequently asked questions

What did Meta actually announce?

Per an internal memo reported by Reuters on 9 July 2026, Meta's in-house AI accelerator, code-named Iris and part of its MTIA family, enters production in September. Broadcom co-designed it and TSMC will make it. At least one chip cleared testing in about six weeks, and Meta aims to roughly double its compute capacity next year from about seven gigawatts this year.

Does this mean Meta is dropping Nvidia?

No. Meta still expects 125 to 145 billion dollars of capital spending this year, much of it on Nvidia and AMD, and has said it will keep buying from both. The in-house chip is a second source and negotiating leverage, not a replacement, which is exactly why it matters for pricing.

Why should a company that does not make chips care?

Because your compute bill does. GPU rental prices carry a premium set largely by one vendor. When the biggest buyers self-supply part of their demand, that premium faces competition, which over time caps how far rental prices can climb. It is a reason to stress-test any 2027 budget built on prices only going up.

A single chip clearing test in six weeks does not end Nvidia's lead, and Meta will keep buying from it for years. What changed is quieter and harder to reverse: the largest buyers now have a credible way to build instead of only buy, and that is the moment a supplier's pricing power stops being a one-way street.

MetaMTIABroadcomTSMCNvidiaAI chipsData centres

Meta Wants to Rent Out Its Spare AI Compute

Bloomberg reports Meta is standing up a cloud business to rent excess AI compute to outsiders, sending its stock past 600 dollars and hitting GPU-rental names. What owners should read into it.

3 min read

Infrastructure

Anthropic Wants Its Own Chip, Not Just Nvidia's

The Information reports Anthropic is in early talks with Samsung to build a custom AI chip. The last big lab that only bought Nvidia is done doing that - here is what it means for the price of compute.

2 min read

Digital Sovereignty

The EU Council Revived the Rule That Scans Private Messages

The EU Council revived the lapsed rule that lets providers scan private messages with AI. For owners, confidential communication is now a governance choice.