If you follow AI infrastructure announcements, you noticed the turn: the fight left the software and went back to the silicon. After two years of a queue at Nvidia, the giants decided to design their own chip for inference in the data center.
In July, Meta confirmed via an internal memo the production of the Iris chip in September. In September, Qualcomm announced a multi-generational agreement with Amazon for custom silicon on AWS. Two different moves, the same thesis: depending on a single supplier had become too expensive and too slow.
This post is an analysis of verified public data, not an account of a bench test. No chip cited was measured here. The numbers come from official statements and from declared dated reproductions, with source, level, and date on each figure, in order to update TOPS and NPU at the edge and AI’s electric wall from the angle that was missing: silicon independence in the data center.
Why make your own chip now
Buying an off-the-shelf GPU worked while the bottleneck was experimenting. At gigawatt scale, three costs became a decision for the board:
- Price and availability. Adopting the newest GPUs at Meta’s scale became, in the words of the internal memo itself, a heavy effort that cost time. Queue, premium, and allocation came to define the roadmap.
- Fit to the real workload. A generic chip serves to train everything. For inference in ranking, recommendation, and agents — the daily bill of Facebook, Instagram, and AWS — a custom-made ASIC delivers more tokens per watt and per dollar.
- A calendar of your own. The typical industry cycle is 12 months or more between generations. Meta plans a new chip every six months through 2027. Having the design in-house is having the launch date in-house.
Making does not mean abandoning Nvidia or AMD. In both cases in this piece, custom silicon complements the volume bought from third parties. It is diversification, not a rupture.
Case 1: Meta Iris — from MTIA to the gigawatt
The chip code-named Iris is one of the four generations of the Meta Training and Inference Accelerator (MTIA) program revealed by Meta in March 2026, then under the technical designations MTIA 300, 400, 450, and 500, with deployment planned over the following 18 months, focused on generative inference, ranking, and recommendation.
What the internal memo reviewed by Reuters on July 9, 2026 adds, in numbers:
- Production in September 2026, at TSMC, with a Broadcom design. The Meta–Broadcom partnership was expanded in early 2026 to co-develop multiple generations of MTIA through 2029.
- A test in six weeks, with no serious problem found. A short pace for a project that, according to the memo itself, had been stalling for more than five years.
- 7 GW of compute in 2026, doubling to 14 GW in 2027. For reference, 1 GW powers about 800,000 homes.
- Infrastructure spending of up to US$ 145 billion in 2026, a relevant share of the more than US$ 700 billion projected for the big tech companies in the year.
- Multi-year contracts locked in: memory with Samsung Electronics, flash with Sandisk, and optical fiber with Sumitomo Electric — precisely the three inputs where scarcity already makes the final product more expensive, as the coverage of the case records.
- A complement, not a swap: Iris is added to the large volumes of GPUs bought from Nvidia and AMD. Meta also maintains a multi-year agreement with AMD for up to 6 GW in Instinct GPUs, according to the same coverage.
The engineering reading is direct: Meta calibrated the chip for its own workload — recommendation and inference at scale — instead of paying for the most expensive universal chip and waiting in its queue.
Case 2: Qualcomm and Amazon — inference and optics on AWS
On September 8, 2026, Qualcomm Technologies announced in an official statement a multi-generational collaboration with Amazon for custom silicon in AWS AI data centers, with a focus on inference, plus high-bandwidth optical connectivity.
The public terms, by layer:
- Technical scope: multiple generations of custom silicon for inference on AWS, plus optical solutions with advanced SerDes and optical DSP, extending to 1.6T and future generations.
- Commercial structure: a multi-generational and multi-product commercial engagement, plus a warrant that gives Amazon the right to acquire a stake of up to about US$ 4 billion in Qualcomm shares.
- Revenue potential: Qualcomm’s CFO and COO, at the Goldman Sachs Communacopia conference on September 8, 2026, spoke of a potential opportunity of up to US$ 60 billion over the contract, with revenue already starting in the December quarter and production underway.
- Engineering counterpart: Qualcomm is expanding its use of AWS AI infrastructure, including Amazon Bedrock, for EDA workloads, aiming to shorten chip-design cycles.
- Market context: Qualcomm’s share rose about 9% to 10% in the premarket of the announcement. It is Qualcomm’s second hyperscaler after the data-center CPU deal with Meta in June 2026, and it comes weeks after a similar agreement between Marvell and Google, with a participation right of up to US$ 12.2 billion.
Here the thesis is the other end of the same coin: the one that dominated the phone chip brings efficiency per watt to the rack, and the one that operates the cloud secures silicon tuned to its own fleet.
Buying versus making: the table that decides
| Criterion | Buy an off-the-shelf GPU | Design your own silicon |
|---|---|---|
| Timeline | Enter the supplier’s queue | Set your own calendar (Meta is aiming for 6 months) |
| Unit cost at scale | Full price + a premium in the shortage | Spread NRE across gigawatts, cheaper per token |
| Efficiency on the real workload | A universal chip, good at everything | An ASIC tuned to inference and recommendation |
| Risk | Dependence on someone else’s roadmap | Years of design, validation, and toolchain |
| Practical outcome | Ideal for experimenting and for peaks | Ideal for a continuous 24/7 base load |
That is why the emerging pattern is hybrid: the state of the art is bought in order to train and absorb peaks, and the daily base runs on the in-house chip.
What this changes for you, who does not buy chips
You are not going to import an Iris or an AWS accelerator. The effect arrives by three paths:
- Price of the cloud and of the device. Memory, flash, and fiber under a long contract hold the cost for those who have scale and pass inflation on to those who do not — the direct link with the rise in laptops and consoles recorded last week.
- Latency and a new feature in the app. Ranking, recommendation, live captions, and semantic search become cheaper per query when they run on a custom ASIC. A feature that was expensive becomes the standard.
- Energy as the limit. An efficient chip does not eliminate the electric wall, but it pushes the frontier: more inference per megawatt postpones the hour when the grid stalls the expansion.
How to read the next announcement without falling for the marketing
Before passing along the press release, apply this four-check routine:
- Require the design–factory pair: who designed it and who manufactures it (e.g., Broadcom + TSMC). An announcement without a foundry is intent, not capacity.
- Separate production from deployment: entering production in September is not being in every data center in September. Look for the target in GW and the revenue quarter.
- Check whether it complements or replaces: if the text says it adds to the Nvidia/AMD volume, read it as diversification. Total replacement in 2026 is a slogan.
- Follow the locked-in input: a contract for memory, flash, and optics is worth as much as the chip. Without a guaranteed input, the roadmap is a letter of intent.
Sources cited and official references
- Official statement from Qualcomm Technologies, Qualcomm Announces Multi-Generational Product Collaboration with Amazon to Build Next-Generation AI Data Center Infrastructure — multi-generational collaboration for inference on AWS and optics up to 1.6T (official statement from the manufacturer, September 8, 2026, verified on 2026-09-13).
- Reuters, Meta to put AI chip into production in September as it looks to double computing capacity, memo shows — Meta internal memo on Iris: production in September at TSMC with a Broadcom design, a six-week test, 7 GW in 2026 and 14 GW in 2027, spending of up to US$ 145 billion (journalistic reproduction of an internal document, July 9, 2026, verified on 2026-09-13).
- CNBC / Data Center Dynamics / TechCrunch, reproductions from July 9–10, 2026 of the same memo — a four-generation MTIA program revealed in March 2026, a six-month cadence through 2027, contracts with Samsung, Sandisk, and Sumitomo Electric (declared dated reproductions, verified on 2026-09-13).
- Meta Newsroom, Meta Partners With Broadcom to Co-Develop Custom AI Silicon and Expanding Meta’s custom silicon to power our AI workloads — partnership expanded through 2029 and four generations of MTIA for ranking, recommendation, and generative inference (official Meta documentation, April and March 2026, verified on 2026-09-13).
- Fierce Network, Qualcomm’s custom-silicon deal with Amazon marks its pivot toward data centers — the CFO’s remarks at Goldman Sachs Communacopia 2026: potential of up to US$ 60 billion, revenue starting in the December quarter, the agreement with Meta in June, and the 200 MW HUMAIN project (sector coverage with an executive statement, September 10, 2026, verified on 2026-09-13).
- Bloomberg / Reuters, coverage from September 8, 2026 of the Qualcomm–Amazon agreement — a warrant of up to about US$ 4 billion in shares and a rise of 9% to 10% in the premarket (declared dated reproductions, verified on 2026-09-13).
