Meta customized AMD MI450 chip’s computing power halved and memory greatly reduced; SemiAnalysis: This is a tragedy of Meta’s corporate culture

Meta's series of missteps in AI infrastructure are exposing deeper organizational cultural issues at the cost of billions of dollars.

According to the latest disclosures from semiconductor industry research firm SemiAnalysis, Meta is requesting AMD to customize a drastically downscaled MI450X chip for them—halving the computing units, reducing HBM memory stacks from 12 layers to 8, and slashing the memory capacity by nearly two-thirds. SemiAnalysis has labeled this decision as "catastrophic," and has publicly called on AMD to bypass Meta's infrastructure team and work directly with Meta's super intelligence laboratory TBD Lab to push for the procurement of the standard MI450X version.

This incident is not an isolated case. SemiAnalysis points out that from the Rivos acquisition worth over $2.5 billion, to the H100 custom server Grand Teton, and then to the GB200 custom solution Ariel, Meta’s infrastructure team has consistently exhibited a pattern of over-engineering, lack of co-design across hardware and software, and short-term political motivations overriding long-term technical rationality. "The Meta infrastructure team is in dire need of a cultural reset," SemiAnalysis writes.

The AMD MI450X is currently one of the most aggressively engineered GPUs in the market: it uses a 2nm process, hybrid bonding packaging technology, 12-layer HBM4 memory stacks, and the largest CoWoS reticle size in the market, representing the cutting edge of packaging and memory density.

However, according to SemiAnalysis, the custom version ordered by Meta will halve the compute die area and reduce HBM stacking from 12-Hi to 8-Hi, significantly compressing both computational power and memory bandwidth. Meta's infrastructure team claims their reasoning is that the configuration is designed for recommendation systems (RecSys) workloads, aiming to increase the CPU-to-GPU compute ratio.

The problem is, this decision was made before the establishment of TBD Lab, which, as Meta's core large model research team, has no interest in this chip. SemiAnalysis makes it clear that compared to Nvidia's Vera Rubin, the stripped-down MI450 holds no appeal for TBD Lab, "TBD will heavily favor Rubin." This means AMD’s shipment volumes to Meta will be severely impacted by this decision.

In a rare move, SemiAnalysis directly urges AMD in their report: "AMD needs to step up and work directly with the TBD team, ensuring they receive the standard MI450, and not this castrated version that is worthless for GenAI."

Another classic case of Meta infrastructure decision errors is the acquisition of Rivos in 2024 for more than $2.5 billion.

According to SemiAnalysis, very few inside Meta's chip division truly understood the strategic logic behind the acquisition, while those who drove the transaction have since fallen silent. The mainstream view is that: with abundant funds, the custom chip sector heating up, and Meta already licensing Rivos IP, management decided it was better to acquire the company outright.

However, the structure of the acquisition itself sowed the seeds of trouble. The founding team of Rivos insisted on selling the entire staff as a package, so Meta had to buy the whole company, and then made massive layoffs of its unwanted departments. Reports cite former Meta chip staff who said the acquisition was led by Meta's silicon head Yee Jiun Song, who faced internal opposition, but lost interest after the deal closed. Current chip team managers viewed Rivos engineers as "free headcount" to be absorbed by their own teams—quickly dismantling the original team structure.

On the technical front, the core value of the acquisition—Rivos's SIMT architecture GPU IP—evaporated after the original "Olympus" chip project that was to use the tech was canceled. The replacement "Phoebe" project is expected to tape out no earlier than 2028, but SemiAnalysis notes internal confidence in its delivery is low.

The staff exodus is just as striking. According to SemiAnalysis, about 30% of Rivos employees have left in recent layoffs, with co-founder Mark Hayter already departed, and several former Rivos team members moving to chip startup Nuvacore founded by Gerard Williams after their first RSUs vested in May 2024. SemiAnalysis also reveals that Rivos CEO and co-founder Puneet Kumar is likely to leave one or two years after his Meta shares fully vest.

Meta’s obsession with customization is hardly new. Its H100 custom server, Grand Teton, adds an extra switch tray on top of the standard HGX server, packing four Broadcom PCIe switches, 16 SSDs, and eight NICs—all to provide each server with more direct-attached storage for training checkpoint storage.

But when put into production, the model teams utilized this extra storage far less than expected, and the design was eventually abandoned. SemiAnalysis points out this is a textbook example of poor coordination between Meta’s hardware and software teams—the infrastructure group paid higher material and energy costs for underused features, and increased reliance on Broadcom, running counter to their original intention to reduce dependence on Nvidia networking equipment.

With the advent of the Blackwell generation, Meta’s custom solution "Ariel" followed a similar logic. The standard GB200 pairs each Grace CPU with two B200 GPUs, but Ariel uses a one-to-one ratio, halving the GPU count—again, supposedly to raise the CPU ratio for RecSys workloads.

According to SemiAnalysis's calculations, the total cost of ownership (TCO) for the Ariel NVL36x2 solution is 14% higher than the standard GB200 NVL72, and the extra CPU/DRAM resources bought with that cost are precisely what the large-model teams do not need. Adopting cross-rack interconnects to compensate for fewer GPUs introduces extra network latency and reliability issues. Meanwhile, the NVL72 backplane—originally deemed unstable—has steadily matured, proving Meta’s risk assessment wrong.

SemiAnalysis notes that Meta's GB300 servers have now reverted to the standard configuration—an implicit rejection of the Ariel approach.

SemiAnalysis attributes these issues to the organizational culture of Meta's infrastructure team.

The firm notes that Meta’s six-month performance cycle eliminates the lowest 10% to 15% of staff in each round, pushing teams to chase quick, visible wins over long-term technical bets. Some managers favor high-profile, quickly deliverable projects and then move on swiftly—a practice called "window washing" internally. Meanwhile, few dare to openly challenge upper management decisions, making it harder to correct mistakes in a timely fashion.

The supply chain team has limited say in engineering matters, aggravating the issues above. SemiAnalysis notes that some suppliers have lowered the priority of Meta’s new projects due to its constant changes of direction, instead prioritizing Amazon or Google needs.

SemiAnalysis compares this phenomenon to the expansion of Meta’s Reality Labs—before mass layoffs, tens of billions of dollars were poured into large engineering teams and R&D projects. Now, as Meta begins to sell compute capacity to external customers, the cost of this internal culture will be more directly exposed to the market.