Meta's Custom AMD Chip Specs Cut in Half; Research Firm Criticizes Its Decision-Making Culture
Alina Collins
Chip research firm SemiAnalysis reports that Meta is asking AMD to halve the compute units and cut memory by nearly two-thirds on a custom MI450X — a decision SemiAnalysis calls 'disastrous' because it could push AMD off the buy list of Meta's new superintelligence lab.
What exactly is Meta asking AMD to cut?
Meta's infrastructure team wants a stripped-down MI450X: compute units halved, HBM memory stacking cut from 12 layers to 8, and memory capacity reduced by nearly two-thirds.
The standard MI450X uses a 2 nm process, hybrid bonding — joining two chips directly face-to-face — and 12-layer HBM4 stacking. It represents the current ceiling for packaging and memory density.
Meta's rationale: the chip targets recommendation workloads and needs a higher CPU-to-GPU compute ratio. This means → Meta decided recommendation systems don't need full-spec compute and set the specs on a "good enough" basis.
Why does SemiAnalysis call this 'disastrous'?
When the decision was made, Meta's superintelligence lab TBD Lab did not yet exist. Now that TBD Lab is online, the cut-down chip holds "no appeal" for it.
SemiAnalysis expects TBD Lab to lean heavily toward Nvidia's Vera Rubin, dealing a serious blow to AMD's actual shipment volume at Meta.
SemiAnalysis went as far as urging AMD to bypass Meta's infrastructure team and engage TBD Lab directly to push full-spec MI450X procurement. In plain terms = the research firm believes Meta's own procurement team is pushing AMD out the door, and AMD should find a different counterpart to talk to.
What did the $2.5 billion+ Rivos acquisition actually deliver?
Meta acquired chip startup Rivos in 2024 for over $2.5 billion, but SemiAnalysis reports that few people inside Meta truly understood the deal's strategic logic. Those who championed it have since gone quiet.
The deal structure itself created problems: Rivos's founding team insisted on an all-or-nothing sale. Meta had to buy the entire company, then laid off the departments it didn't need. Existing team managers treated Rivos engineers as "free headcount," and the original team structure collapsed.
The acquisition's core value — Rivos's SIMT-architecture GPU intellectual property — vanished when the "Olympus" chip project was cancelled. The replacement, "Phoebe," is not expected to tape out before 2028 at the earliest, and internal confidence is low.
Talent attrition has been severe: roughly 30% of Rivos employees left in recent layoffs, co-founder Mark Hayter departed, and several former Rivos staff joined chip startup Nuvacore after their first restricted stock units vested in May.
Grand Teton and Ariel — how much has the customization obsession cost?
Meta's custom H100 server, Grand Teton, added a switch tray, SSDs, and NICs to give each server more direct-attached storage for training checkpoints. In production, model teams used far less storage than expected, and the design was abandoned.
For the Blackwell generation, Meta's custom design Ariel changed the standard GB200's two-B200-GPUs-per-Grace-CPU ratio to one-to-one, halving GPU count. SemiAnalysis estimates Ariel's total cost of ownership runs 14% higher than the standard configuration — the extra spending bought CPU and memory resources that large-model teams do not need.
Meta's GB300 servers have reverted to the standard configuration. This means → Meta's own actions repudiated the Ariel design, but the cost of the detour has already been paid.
Where does the cultural root lie?
SemiAnalysis traces the problem to Meta's infrastructure team and its six-month performance-review cycle, which culls the bottom 10%–15% each round. Teams gravitate toward short-term visible wins over long-term technical planning.
Some managers pursue high-visibility, quick-ship projects and move on immediately after delivery — a practice known internally as "window washing." Supply-chain teams have limited voice in engineering decisions, amplifying the problem.
This reflects a deeper signal: some suppliers have already deprioritized Meta's new projects because of its frequent direction changes, choosing to serve Amazon or Google first.
As Meta moves to sell compute to external customers, the cost of this internal culture will be exposed directly to the market — at that point, the price of poor decisions will no longer be absorbed internally.
Content is for reference only, not financial advice.