Meta's Bet to Fold Nvidia's Invoice in Half

Here's the deal: Meta just crossed the line from a company that only buys other people's chips to one that stamps out its own. According to an internal memo reviewed by Reuters, Meta will begin mass production of its in-house data center AI chip, code-named "Iris," starting in September. Testing wrapped in just six weeks, and the memo says it turned up no major issues. That's a bigger signal than it sounds — a brand-new piece of silicon passing validation cleanly in six weeks means the design came out remarkably stable.

But the real story isn't one chip. Meta is using Iris as the foundation to double its total compute capacity from 7 gigawatts (GW) in 2026 to a full 14GW in 2027. To lay that infrastructure, it plans to spend as much as $145 billion this year alone. For context, that's a big slice of the more than $700 billion Big Tech is collectively expected to pour into AI in 2026 — carried by Meta by itself.

Pulling off a build this large takes more than a chip. You have to lock down memory, NAND flash, and even the fiber-optic cabling in advance. So Meta signed long-term supply deals with Samsung Electronics (memory), SanDisk (flash storage), and Sumitomo Electric (fiber optics). The market reacted instantly. SanDisk jumped as much as roughly 7% intraday, Sumitomo Electric's US-listed ADR rose around 4.7%, and Samsung gained about 2.5% in Seoul. The main character, Meta itself, actually dropped around 4% before bouncing back to close roughly flat — and there's a reason for that. Let's unpack it piece by piece.

Meet the Players

Meta runs Facebook, Instagram, and WhatsApp, but over the past few years its identity has been shifting fast toward "AI infrastructure company." The catch is that most of the compute powering that AI leans on Nvidia and AMD GPUs. GPUs are expensive, perpetually scarce, and Nvidia holds the pricing power. From Meta's seat, that's a structure where it hands tens of billions a year to someone else. Iris is the attempt to break it.

Iris is the fourth generation of Meta's in-house accelerator program, MTIA (Meta Training and Inference Accelerators). As the name says, it targets both training and inference. The crucial detail is that Meta frames it as a complement, not a replacement, for Nvidia. The goal isn't to rip out GPUs entirely — it's to run Meta-specific workloads, like the recommendation engines behind Facebook and Instagram, on custom silicon and shave the overall bill.

Broadcom is the design partner behind Iris. When it comes to turning a customer's target spec into an actual silicon design, Broadcom is one of the industry's heavyweights. TSMC handles the foundry work that actually fabricates that design. So the division of labor is a triangle: Meta specs it, Broadcom designs it, TSMC builds it. That combination is a near carbon copy of the path Google walked to build its own TPU.

Don't forget the supply-chain protagonists either. Samsung Electronics provides the high-bandwidth memory (HBM) and general memory to fill these data centers, while SanDisk supplies the NAND flash to store vast piles of AI data. Worth noting: SanDisk spun off from Western Digital and listed independently in 2025, so a single large contract like this shows up sharply in its stock. Sumitomo Electric supplies the fiber-optic cabling that links everything inside and between data centers. No matter how fast the compute is, if the plumbing that moves data clogs up, it's useless — fiber is the quiet but essential part.

What the Memo Actually Says

The root of this reporting is an internal memo. Meta didn't put out a press release; Reuters obtained a company memo and broke it as an exclusive, and CNBC picked it up. So it's worth keeping in mind this isn't an official, company-confirmed announcement. That said, when outlets at the level of Reuters and CNBC state they've seen the physical memo, the reliability is high.

The memo boils down to three chunks. First, the chip: Iris goes into mass production in September, having cleared testing in six weeks with no defects. Second, compute expansion: 7GW in 2026 doubling to 14GW in 2027. Third, the money: up to $145 billion on AI infrastructure this year. The three interlock — source your own chips to cut unit costs, use those savings to lay more GW, and lock down the supply chain in advance to do it.

Here's the whole thing at a glance.

Item Detail
Chip code name Iris (MTIA gen 4)
Mass production start September 2026
Testing period ~6 weeks, no major issues
Design partner Broadcom
Manufacturing TSMC
2026 compute capacity ~7GW
2027 target capacity 14GW (double)
2026 AI infra capex Up to $145 billion
Supply deals Samsung (memory), SanDisk (flash), Sumitomo Electric (fiber)

Now the stock reaction. On July 9, the day the memo surfaced, the market graded the component suppliers and Meta itself in opposite directions. SanDisk surged as much as ~7% intraday (some tallies put it into double digits), Sumitomo's ADR rose 4.7%, and Samsung gained about 2.5% in Seoul. Storage names like Micron, Western Digital, and Seagate rallied alongside, lifting the whole sector.

Meta, by contrast, dropped about 4% early in the session before rebounding to roughly flat by late morning. Why did only the main character fall? The $145 billion capex figure spooked the market. That number exceeds Meta's expected operating cash flow for the year (around $136.6 billion). Put simply, it's a plan to spend more than it makes — which triggered a flash of "will this investment actually pay off?" anxiety that pressed the stock down. Then the story of cutting costs with in-house silicon got re-rated, and it bounced.

Who Gains What

Meta gains two things: cost and leverage. With a single Nvidia GPU running into the tens of thousands of dollars, a company like Meta — with clear, repetitive workloads like recommendation and ranking — can slash unit costs with a chip built just for those jobs. On top of that, the mere fact of having its own chip becomes a bargaining chip against Nvidia and AMD. Even without full replacement, the credible possibility of replacement changes the negotiating table.

Samsung Electronics locks in long-term demand for memory, the core of any AI data center. HBM and other high-performance memory are simultaneously the biggest bottleneck and the biggest profit engine of the current AI boom. Pinning a long-term deal with a mega-customer like Meta secures stable volume and pricing while building a valuable reference in its HBM rivalry with SK Hynix.

SanDisk stamps its presence with this one. As a newly independent, spun-off company, landing large customers is its survival and its growth, so a multi-year deal with an anchor customer the size of Meta was a strong enough signal to move the stock immediately. It rides straight into the trend of AI data centers demanding huge volumes of storage for training and inference.

Broadcom and TSMC are already the biggest winners of the custom AI-chip era. Every time a Big Tech firm decides to roll its own silicon, the design and manufacturing orders flow to these two. Iris just stacks another large Meta job onto that pipeline. Sumitomo Electric is quieter, but as data centers multiply, fiber volume mechanically follows — making it a steady beneficiary.

Precedents — the Wins and the Flops

The original success story of the in-house chip gambit is Google's TPU. Google unveiled its own AI accelerator in 2016 and, generation after generation, embedded it deep into its search, translation, and cloud. It's now mature enough to sell to outside customers via the cloud. The key to that win was "you know your own workload best." Because Google knew exactly what it needed to run, it could build a chip optimized for it — and Meta's Iris stands on exactly the same logic.

Amazon went the same route with its Trainium and Inferentia chips and had reasonable success trimming its Nvidia dependence. What both companies share is that they didn't overreach by trying to eliminate Nvidia entirely. They kept buying GPUs but moved only standardized, high-volume workloads to their own chips — a pragmatic strategy to cut costs. Meta's repeated insistence that Iris is a "complement, not a replacement" shows it learned that lesson precisely.

There's a warning bell too. In-house chips require enormous upfront investment in design, validation, and mass production, and if you can't build the software ecosystem (something like CUDA), developers won't use your hardware no matter how good it is. That's why several startups and even some large firms quietly shelved chip projects that performed fine but had "nobody to use them." Meta's edge is that it has massive guaranteed internal demand in its own apps — if you build it, there's definitely somewhere to use it — which sharply reduces that risk.

And the $145 billion capex is itself a risk. Just as telecoms once ached from over-investing during the fiber-optic boom, today's AI infrastructure race stands in front of the same question: "will demand really show up at this scale?" Meta's stock dropping first on announcement day is proof the market hasn't forgotten that question.

How Rivals Counter

The first to react is obviously Nvidia. If one of its biggest customers, Meta, raises the share of its own chips, GPU orders could shrink over the long run. That said, Nvidia still dominates on the newest training GPUs, and its CUDA ecosystem is a deep moat. Nvidia's counter will be to hold the "the heaviest frontier training is still ours" position while strengthening lower-cost inference lineups and software lock-in.

AMD is in a tighter spot. Meta has bought a fair amount of AMD GPUs, and Iris could eat into that volume. AMD has to wedge into the middle ground — "cheaper than Nvidia, more general-purpose than your own chip" — using price-performance and its open software stack (ROCm) as weapons.

Google and Amazon might actually welcome this news. The more the in-house chip route becomes an industry standard, the more their multi-generation head start reads as a relative advantage. They'll lean on their maturity — "we already sell to outside customers" — to differentiate in the cloud race.

The Samsung vs SK Hynix dynamic is fascinating too. Samsung grabbed one axis of Meta's memory supply this round, but SK Hynix, the HBM market leader, won't sit still. Meta's 14GW expansion is an event that grows the entire memory pie, opening a bigger slice for both while making the share fight inside it fiercer. Storage rivals like Western Digital and Micron will watch SanDisk's Meta win and scramble to grab their share of AI storage demand.

So What Actually Changes

For developers — not much right now. Iris is Meta-internal silicon, not something exposed via API. But the trend is worth reading. As Big Tech each moves to its own chips, AI infrastructure fractures from an "Nvidia single standard" into a "multi-pole regime of company-specific custom silicon." Portable code and abstraction layers that don't lock you to specific hardware will only grow more valuable.

For investors — the winners of this news were clear. Component suppliers (SanDisk, Samsung, Sumitomo) rose immediately, while Meta dropped first under capex weight before rebounding. Two things to watch: one, whether the $145 billion actually gets recouped in revenue and cost savings; two, how much the spread of in-house chips erodes the long-term growth picture for Nvidia and AMD. But don't call it too early — remember this is based on an internal memo, not an official announcement.

For enterprises — if you run high-volume, standardized AI workloads, the "own chip vs off-the-shelf GPU" break-even math becomes an increasingly real topic. Not everyone can stamp out chips with Broadcom and TSMC, of course. But once a mega-player like Meta paves the road, intermediaries who lease custom silicon as a service may proliferate and lower the barrier.

For everyday users — Facebook and Instagram recommendations could get smoother, and AI features could run cheaper and faster. If Meta lowers its infrastructure costs, it can use that headroom to layer more AI features onto free services. It's not a chip you'll ever see, but a change that quietly seeps into your feed experience.

🥄 Three Things You're Probably Wondering

— So what does this mean for me? Almost no direct impact. But if you hold semiconductor stocks like Samsung, SanDisk, or Nvidia, or use Facebook and Instagram daily, you're indirectly connected. It's a tailwind for the suppliers, and for Facebook users it can come back as "cheaper, faster AI features."

— So Meta stops buying Nvidia now? No, that's a misread. Meta explicitly frames Iris as a complement to, not a replacement for, Nvidia and AMD GPUs. The heaviest training still runs on GPUs; only standardized, repetitive workloads move to the in-house chip to cut costs. Nvidia orders aren't vanishing anytime soon.

— Is this a confirmed announcement? That's the subtle part. Meta didn't issue a press release — Reuters obtained an internal memo and broke it as an exclusive. Since a high-reliability outlet says it saw the physical memo, it carries weight, but bear in mind it isn't a company-confirmed announcement.

References

Numbers and criteria are as of announcement and may change. Investment calls are yours to make!