Each time you scroll a social feed, software picks the next post, video and ad from a huge pool of options. It does that for billions of people at once. The work runs on special chips packed into large data centers, and the bill for those chips has become a major cost for the companies that run them.
Most of those chips come from a few makers. Many of them are graphics processing units, or GPUs, built to handle almost any AI task. A buyer pays for that range and for the maker's profit, and when demand runs ahead of supply, it waits in the same queue as every other buyer.
So some big buyers have started to design chips of their own. They hire a design partner, send the plans to a contract factory and shape the chip around the few tasks they run all day. They keep buying standard GPUs for the rest.
Meta Platforms, the company behind Facebook, Instagram and WhatsApp, does both. It designs a chip family called the Meta Training and Inference Accelerator (MTIA) and buys the rest from Nvidia, AMD and others. On BloFin you can follow Meta's share price by holding tokenized Meta (METAX), a token that tracks the share, on METAX/USDT Spot, or take a leveraged long or short through the METAUSDT Perpetual.
What is Meta's MTIA chip?
Meta's MTIA chips are the AI chips Meta designs for its own data centers, built around the work its apps do all day. On March 11, 2026, Meta said it runs hundreds of thousands of MTIA chips for inference on posts and ads across its apps. It plans four new generations within two years (source: Meta's March chip post).
An accelerator is a chip designed for one kind of math, the math AI models run on, done very fast. It sits next to a server's main processor, which hands it the heavy lifting. A GPU is also an accelerator, but it is sold to everyone and built to cope with any model. MTIA is shaped around the models Meta runs.
AI work comes in two stages, and the difference explains most of Meta's plan. Training is when a model learns from huge amounts of data, which can take weeks on thousands of chips. Inference is when the finished model does its work, such as picking your next Reels video or the ads Meta sells beside it. That runs every time someone opens one of Meta's apps.
Meta started with chips for its ranking systems, the software that decides what shows up in your feed and which ads you see. It is now moving MTIA into generative AI, the models that write text or make images. Every one of these chips is paid for out of Meta's AI capital spending, the money it puts into servers, buildings and network gear each year.
MTIA generations from MTIA 300 to MTIA 500
Meta's current MTIA roadmap has four chips. MTIA 300 is in production, training the systems that rank posts and ads. Meta says MTIA 400, 450 and 500 will be able to run any of its AI work. Their main use will be generative AI such as Meta's Muse AI in the near future and into 2027 (source: Meta's March chip post).
Chip | Main work, as Meta describes it | Status in Meta's March 2026 plan |
|---|---|---|
MTIA 300 | training ranking and recommendation models | in production |
MTIA 400 | any AI work, mainly generative AI inference | planned for the near future and into 2027 |
MTIA 450 | generative AI inference first, plus ranking work | planned for the near future and into 2027 |
MTIA 500 | generative AI inference first, plus ranking work | planned for the near future and into 2027 |
Meta says each generation brings large gains in computing power, memory bandwidth and efficiency, and that it can release a new chip every six months or less (source: Meta's March chip post). Memory bandwidth is the speed at which a chip reads data from its own memory, and AI models are often held back by exactly that.
More detail on MTIA 400 came in August at Hot Chips, a yearly chip-design conference. The Register reported on August 26, 2026 that the chip carries 288 GB of fast memory in eight stacks of 36 GB. It also said 72 of the chips work together in one rack (source: The Register on MTIA 400).
By The Register's account, MTIA 400 runs about 20% faster than Nvidia's top Blackwell chips at the number formats used for training, at about the same power. It is also about three times slower than Nvidia's newer Rubin chip. The report's view was that Meta's largest model training most likely still runs on Nvidia and AMD GPUs (source: The Register on MTIA 400).
The Iris chip in Meta's July memo
In July, Reuters reported on an internal Meta memo about a chip code-named Iris, part of the four-generation MTIA series. The memo said making the chip would begin in September, after six weeks of testing that found no major issues. Reuters added that Meta plans a new chip about every six months through 2027, and Meta declined to comment (source: Reuters via The Standard).
Why Meta designs chips as well as buying them
Meta builds MTIA to run its AI work for less than standard chips cost and to lean less on any one seller. Meta says mainstream chips are built for the hardest task, training the largest models, and are then used less cost-effectively for inference, the work Meta's apps do for every user (source: Meta's March chip post).
Cost is the plainest reason, and the seller's margin shows why. Gross margin is the share of each sales dollar left after the cost of making the product. Nvidia's was 75.0% in its quarter to July 26, 2026 (source: Nvidia's Q2 fiscal 2027 results).
Say a buyer spends $10 billion on chips from a seller with that margin. About $7.5 billion of the bill is the seller's gross profit and about $2.5 billion covers making the chips. A custom chip still pays a design partner and a factory, so part of that margin stays with them. On the work its own chip can take over, though, the buyer keeps some of what would have gone to the GPU maker.
Bargaining power is the second reason. On Meta's July 29, 2026 results call, finance chief Susan Li said the custom silicon effort "will provide long-term strategic flexibility and supply chain leverage" (source: Meta's Q2 2026 call). A buyer that makes some of its own chips can press harder on price and delivery dates with outside sellers.
Speed is the third reason, and the memo Reuters reported said that adopting the newest GPUs "has been a heavy lift, and it has cost us time" (source: Reuters via The Standard). Meta's own work keeps growing and changing, in its words, and a new chip every six months or less lets the design keep up with its models (source: Meta's March chip post).
The bitcoin mining parallel
Crypto holders have seen this shift before, in bitcoin mining. Mining began on ordinary computer processors and moved to graphics cards. It changed again in 2013, when the first chips made only for bitcoin mining came out. Each was an application-specific integrated circuit, or ASIC, built to do that one task (source: Bitcoin Wiki on mining). The new chips did far more work for the same power, and GPU mining stopped paying.
The parallel holds where a task is fixed and huge. Bitcoin mining repeats one calculation, called hashing, over and over, and ranking posts and ads is a steady, huge workload for Meta. Training the newest AI models keeps changing, and that is the work where Meta still relies on GPUs from Nvidia and AMD.
Broadcom, TSMC and the partners behind MTIA
Broadcom helps design MTIA chips, and TSMC manufactures them. On April 14, 2026, Meta extended its Broadcom deal to cover several more MTIA generations. The first phase commits more than 1 gigawatt of computing, which Meta called the start of a multi-gigawatt rollout (source: Meta's Broadcom post).
A gigawatt here measures power, the electricity a data center full of chips draws, and Meta, AMD and Broadcom all size these deals that way. The deal runs through 2029, and Broadcom said the new MTIA chips will be the first custom AI chips made on a 2-nanometer process (source: TNW on the Broadcom deal). A smaller process fits more transistors on each chip, which usually brings more speed for the same power.
Broadcom brings more than design help. Its Ethernet network gear links Meta's AI clusters, the large groups of chips that work on one task together (source: Meta's Broadcom post). TSMC, the contract chip factory in Taiwan, makes the chips from that design, Reuters reported (source: Reuters via The Standard).
The deal also changed Meta's boardroom, because Broadcom's chief executive, Hock Tan, sat on Meta's board. Given the size of the wider deal, Meta said he would leave the board and move to an adviser role (source: Meta's Broadcom post). TNW reported that he would step off when his term ended at Meta's next annual meeting, then advise on Meta's custom chip strategy (source: TNW on the Broadcom deal). That meeting was held on May 27, 2026, and Tan was not among the twelve directors elected (source: Meta 8-K on the 2026 annual meeting).
Meta's MTIA network news from October 2026
On October 12, 2026, at the Open Compute Project summit, Meta described Tahan, a network switch it built for MTIA 300 chips. Tahan links the chips inside a rack over standard Ethernet and gives each chip 6.4 terabits a second. It is already shipping (source: Meta's OCP post).
The Open Compute Project is a group where big data-center owners share hardware designs, and its yearly summit is where Meta shows its own. Chips that train one model together swap data all the time, so the links between them matter almost as much as the chips. Meta calls this the scale-up network: the fast, low-delay links between the chips inside a rack.
Four Tahan switches connect 16 MTIA 300 chips, and at 6.4 terabits a second each, those 16 chips share 102.4 terabits a second. That matches half the capacity of the four 51.2-terabit switches, the share Meta sets aside for linking the chips. Tahan is built on Broadcom's Tomahawk 5 switch chip (source: Meta's OCP post).
Meta also previewed the next step, a rack design that joins 72 accelerators using 102.4-terabit Broadcom switch chips. Some of the network comes from elsewhere, though. One version of Meta's new IceCube switch blade uses an Nvidia chip, part of Meta's growing ties to Nvidia that now reach the network as well as the GPUs (source: Meta's OCP post).
Who supplies Meta's other AI chips
Nvidia and AMD sell Meta the GPUs that work beside MTIA. Nvidia's multiyear deal, announced on February 17, 2026, covers millions of its Blackwell and Rubin GPUs. A week later, on February 24, Meta signed up for as much as 6 gigawatts of AMD's Instinct GPUs (source: Nvidia newsroom) (source: Meta on the AMD deal).
Supplier | What it supplies Meta | First reported or announced | Confirmed by |
|---|---|---|---|
TSMC | manufacturing of MTIA chips | May 18, 2023 (first MTIA chip) | Meta; current series reported by Reuters |
Nvidia | millions of Blackwell and Rubin GPUs | February 17, 2026 | Nvidia |
AMD | up to 6 gigawatts of Instinct GPUs | February 24, 2026 | Meta and AMD |
TPU chips rented through Google Cloud; purchase still in talks | February 26, 2026 (rental) | reported by The Information | |
Broadcom | design work on MTIA chips, plus Ethernet networking | April 14, 2026 (extension) | Meta |
Meta presents the AMD deal as part of a "portfolio-based approach", and Mark Zuckerberg called it "an important step for Meta as we diversify our compute". The first shipments begin in the second half of 2026 (source: Meta on the AMD deal). AMD's side, a version of its MI450 chip tuned for Meta and a warrant that lets Meta buy AMD shares, is set out in the AMD-Meta deal explained.
Google's role is less settled. In February 2026, The Information reported that Meta had agreed to a multibillion-dollar deal to rent Google's own AI chips, called TPUs, through Google Cloud. The same report said Meta was also talking to Google about buying TPUs for its own data centers, and that those talks had not produced an agreement (source: SiliconANGLE on the Google deal).
Gigawatts make the deals easier to compare. The memo Reuters reported put Meta's plan at 7 gigawatts of computing in 2026, doubling in 2027 (source: Reuters via The Standard). Broadcom's first phase, at more than 1 gigawatt, is more than a seventh of that 2026 figure, since 1 gigawatt is 14.3% of 7. AMD's deal, at up to 6 gigawatts, equals 85.7% of it, though the deal runs over several years and its shipments only begin in the second half of 2026.
Buying chip teams: Rivos and FuriosaAI
Meta has also bought chip know-how outright. In October 2025, Walden Catalyst, an investment arm of Rivos backer Walden International, confirmed Meta's purchase of Rivos, a startup that designs chips on RISC-V, an open standard. Rivos had already been helping Meta design its MTIA chips (source: The Next Platform on the Rivos deal).
RISC-V, named after reduced instruction set computer, is a free and open set of basic commands a processor understands. A company can build on it without paying another chip designer for a license. Meta did not disclose the price. In August 2025, The Information had reported that Rivos was seeking funding that would value it above $2 billion (source: The Next Platform on the Rivos deal).
An earlier attempt had failed. In March 2025, FuriosaAI, a South Korean AI chip startup, reportedly turned down an $800 million offer from Meta. The talks broke down over plans for the business and how it would be run after a sale, and price was not the issue, according to the local media report TechCrunch cited (source: TechCrunch on FuriosaAI).
Both moves point at the same scarce input, the engineers who can design AI chips. A company that wants a new chip every six months needs more of them than it can hire one by one. Buying a startup brings in a whole team at once.
What chip choices do to Meta's capex bill
Every chip route lands somewhere in Meta's accounts. Chips Meta buys, or has made for its own data centers, count as capital spending. Meta expects that to total $130 billion to $145 billion in 2026, after $31.1 billion in the April-to-June quarter alone (source: Meta's Q2 2026 results).
That one quarter came to between 21.4% and 23.9% of the full-year range. Free cash flow, the cash left after that spending, fell to $784 million in the quarter from $8.5 billion a year earlier (source: Meta's Q2 2026 results). The cost of a chip Meta owns is then charged against profit a slice at a time, as depreciation, over the years the server is used.
The chip choice changes the price of each gigawatt, but it leaves the number of gigawatts Meta wants the same. If MTIA does the same work for less, each new rack costs Meta less, and the same budget stretches further. If a generation slips or falls short of its plan, Meta fills the gap with GPUs bought at the seller's price, and the bill rises.
Demand is the other half of the bill. On the July call, Zuckerberg said a substantial amount of Meta's compute goes to training models, the work of Meta Superintelligence Labs. Li added that Meta was not yet giving a figure for its 2027 capital spending (source: Meta's Q2 2026 call). The chips also need buildings and power, which is where the data centers Meta is financing with partners come in.
Chip questions in Meta's third-quarter report
Meta updates its spending plans with each quarterly report, and it set its third-quarter 2026 results for after Nasdaq closes on Wednesday, October 21, 2026 (source: Meta Investor Relations). That report can answer four questions about chips. The first is whether Meta keeps its 2026 spending range of $130 billion to $145 billion.
The second is whether Meta gives a first view of 2027 spending, which it held back in July. The third is whether it says anything new about its own chips. The fourth is whether Meta reports progress on MTIA 400, the next chip in its March roadmap.
Each answer speaks to a different part of the bill. A higher range means more chips and buildings this year, and a 2027 figure shows how long the build lasts. Progress on MTIA 400 tells you how much of the work Meta can move onto its own chips.
Meta releases results after the close, and METAX keeps trading on its weekday schedule while Nasdaq is shut, so its first move after the release shows up on the METAX/USDT Spot page. The METAUSDT Perpetual page lists the contract's mark price and funding, the fee one side of the contract pays the other.
What Meta's chip plans mean for METAX and METAUSDT holders
MTIA reaches a METAX holder through cost. Tokenized Meta (METAX) tracks Meta's share price, so the chip plan shows up for you in Meta's spending, profit and free cash flow. A plan that lowers the cost of each gigawatt leaves more free cash flow from the same revenue. That line has shrunk sharply as the build has grown.
The share's move on any one night depends on the whole report, with ad sales and costs weighed alongside the chip plan. What the chip story gives you is a way to read the spending lines when they arrive. You can judge whether a change in spending comes from more chips, more buildings or a new supplier mix.
A leveraged position stacks extra risk onto the share's normal swings. A spending figure that surprises the market can send Meta's price sharply up or down after the close, and leverage multiplies that move against your margin. Size a METAUSDT Perpetual position the way you would when trading Meta around earnings, with room for a gap after the close. Outside the Nasdaq session the perpetual may accept only orders that reduce or close a position, so plan any additions before the close (source: BloFin stock futures notice).
Looking to trade Meta as its chip plans unfold? To get started, you'll need to first create a BloFin account, fund your account with cryptocurrency, and go to the METAX/USDT Spot trading page or the METAUSDT Perpetual trading page.
Frequently asked questions
What software runs on Meta's MTIA chips?
MTIA runs on open software that many AI developers already use. Meta builds its MTIA software around PyTorch, a widely used AI framework. It adds vLLM, a tool for serving language models, and Triton, a language for writing fast chip code. Meta says building on these common tools is meant to make adopting MTIA chips frictionless, so models its teams already wrote can move onto MTIA with little extra work (source: Meta's March chip post).
What is the difference between a GPU and an ASIC?
A GPU is a general-purpose accelerator first built for graphics, and it can run almost any AI model. An ASIC is designed for one kind of task, so it does that task faster or with less power but adapts poorly to others. Meta's MTIA chips sit in the ASIC family, since Meta shapes them around its own ranking and inference work.
How does MTIA 300 compare with a GPU on cost?
At the Hot Chips conference in August 2026, Meta's presenters said MTIA 300 reaches parity with GPUs at a competitive total cost of ownership, the full cost of buying, powering and running a chip over its life. They also reported a speedup of more than 1.8 times on the main training steps, and the chip pairs with 216 GB of fast memory (source: ServeTheHome on MTIA at Hot Chips). These are Meta's own figures, with no dollar amounts given.
Which companies supply memory for Meta's AI build?
According to the memo Reuters reported in July 2026, Meta signed long-term, multi-year supply deals with Samsung for memory chips and Sandisk for flash storage during a shortage of memory chips. It also signed one with Sumitomo Electric for fiber-optic equipment (source: Reuters via The Standard). AI chips need large amounts of fast memory beside them, so memory supply affects how many chips Meta can put to work.
How does MTIA 300 connect to other chips?
MTIA 300 carries its network connections inside the chip package. Two network chiplets, each with six 800-gigabit network interface controllers, or NICs, give the chip 1.2 terabytes a second of total input and output. Data skips the trip through the server's main processor. On a large recommendation model spread across 40 chips, Meta says this made communication between chips 3.9 times faster than on a matching GPU cluster (source: Meta's MTIA 300 post).
Researched and written by the BloFin Academy editorial team with AI-assisted drafting. All facts independently verified. Primary sources include Meta's March 2026 custom silicon post, its Broadcom and AMD announcements, its engineering posts on MTIA 300 and the OCP summit, its second-quarter 2026 results, call and third-quarter results notice, its 8-K on the 2026 annual meeting, Nvidia's announcement of the Meta partnership and its second-quarter fiscal 2027 results, and the Bitcoin Wiki, plus BloFin's stock futures notice. News sources include Reuters, The Register, ServeTheHome, TNW, The Next Platform, TechCrunch, SiliconANGLE and The Information, current as of October 21, 2026.
Nothing in this article constitutes financial advice, and nothing in it is a recommendation to buy, sell or hold Meta in any form. Meta's chip plans run on dates it sets for itself, and a delayed or weaker MTIA generation would leave it buying more GPUs at outside prices. METAX follows any move in Meta's shares after a spending update, and a leveraged METAUSDT Perpetual position can lose its whole margin if the price gaps past its liquidation level. Past performance does not indicate future results. Do your own research and consider your risk tolerance before you trade on BloFin.