The Custom Chip: A "Half-MI450" Built for Recommender Systems

The custom MI450-based accelerator is not a product available on the open market. It is a co-engineered variant purpose-built for Meta's internal workloads. According to SemiAnalysis, the chip uses AMD's chiplet architecture to create a version with significantly reduced compute and memory resources, tuned specifically for the memory-bandwidth-per-dollar ratio that recommendation systems demand .

Meta's reasoning, as reported, is to optimize CPU-to-GPU compute ratios at the rack level and improve memory bandwidth cost-efficiency for its RecSys (recommendation system) workloads, which run billions of inferences per day .

Compute Dies and Memory: How the Custom Chip Differs from the Standard MI455X

The standard Instinct MI455X, which powers AMD's Helios rack-scale system, is an engineering flagship built on the CDNA 5 architecture. It uses 8 compute chiplets on TSMC's N2 (2nm) process and 12 HBM4 12-Hi stacks to deliver 432 GB of memory . Meta's custom version cuts the compute dies in half—down to 4—and drops to 6 HBM4 8-Hi stacks, yielding only 144 GB of memory . This is essentially half the compute silicon and one-third of the memory bandwidth of the fully-fledged MI455X.

Because the chip uses fewer chiplets and HBM stacks, Meta's variant is also physically smaller and uses less power and packaging complexity, reducing cost per accelerator .

SemiAnalysis Criticisms: A "Catastrophic" Blunder or Pragmatic Engineering?

SemiAnalysis levied several sharp criticisms at Meta's decision:

• "Catastrophic" blunder in AI infrastructure: The firm argued that Meta forced AMD to "cripple its most powerful chip," calling the move short-sighted .
• Cultural dysfunction: SemiAnalysis framed the decision as a symptom of Meta's "performance-driven culture run amok," where excessive customization stifles long-term hardware competitiveness .
• Public call to bypass Meta's infrastructure team: SemiAnalysis openly urged AMD to circumvent Meta's internal infrastructure group and instead work directly with Meta's "TBD Lab" super-intelligence unit, which would presumably opt for the full MI455X .
• Micromanagement and repeated missteps: They pointed to a pattern of "catastrophic" infrastructure decisions, including a separate $2.5 billion acquisition that "ended in utter chaos" .

The core complaint is that by stripping down the chip, Meta sacrifices versatility and future-proofing for narrow near-term cost optimization, potentially locking itself out of more capable hardware needed for frontier AI workloads .

Not all observers agree. Some defense of the approach notes that recommendation systems do not need frontier LLM capacity. What they need is bandwidth per dollar on a task run billions of times a day, making the cut-down chip a rational specialization rather than a mistake .

Fit Within Meta's Five-Year, $60 Billion AMD Partnership

On February 24, 2026, Meta and AMD announced a five-year, $60 billion strategic partnership to deploy up to 6 gigawatts of AMD Instinct GPUs for Meta's AI infrastructure . Key terms include:

• Scope: AMD will supply custom MI450-based accelerators (the cut-down variant above), plus two generations of Epyc CPUs, across Meta's global data centers .
• Performance-based warrant: AMD issued Meta a warrant for up to 160 million shares of AMD common stock, at an exercise price of $0.01 per share. This represents approximately 10% of AMD's outstanding shares .
• Vesting schedule: The warrant vests in tranches tied to deployment milestones. The first tranche vests upon the initial 1 GW of shipments, with additional tranches unlocking as Meta scales purchases toward the full 6 GW .
• Price hurdles: Further vesting is tied to AMD's stock achieving price thresholds up to $600 per share .

The custom "Half-MI450" is the first concrete product manifestation of this megadeal: a chip tailored specifically to Meta's recommendation inference needs, designed to maximize cost- and power-efficiency per inference at massive rack-level scale, exactly what the $60 billion partnership was structured to enable .

The Custom Chip: A "Half-MI450" Built for Recommender Systems

The custom MI450-based accelerator is not a product available on the open market. It is a co-engineered variant purpose-built for Meta's internal workloads. According to SemiAnalysis, the chip uses AMD's chiplet architecture to create a version with significantly reduced compute and memory resources, tuned specifically for the memory-bandwidth-per-dollar ratio that recommendation systems demand .

Meta's reasoning, as reported, is to optimize CPU-to-GPU compute ratios at the rack level and improve memory bandwidth cost-efficiency for its RecSys (recommendation system) workloads, which run billions of inferences per day .

Compute Dies and Memory: How the Custom Chip Differs from the Standard MI455X

The standard Instinct MI455X, which powers AMD's Helios rack-scale system, is an engineering flagship built on the CDNA 5 architecture. It uses 8 compute chiplets on TSMC's N2 (2nm) process and 12 HBM4 12-Hi stacks to deliver 432 GB of memory . Meta's custom version cuts the compute dies in half—down to 4—and drops to 6 HBM4 8-Hi stacks, yielding only 144 GB of memory . This is essentially half the compute silicon and one-third of the memory bandwidth of the fully-fledged MI455X.

Because the chip uses fewer chiplets and HBM stacks, Meta's variant is also physically smaller and uses less power and packaging complexity, reducing cost per accelerator .

SemiAnalysis Criticisms: A "Catastrophic" Blunder or Pragmatic Engineering?

SemiAnalysis levied several sharp criticisms at Meta's decision:

• "Catastrophic" blunder in AI infrastructure: The firm argued that Meta forced AMD to "cripple its most powerful chip," calling the move short-sighted .
• Cultural dysfunction: SemiAnalysis framed the decision as a symptom of Meta's "performance-driven culture run amok," where excessive customization stifles long-term hardware competitiveness .
• Public call to bypass Meta's infrastructure team: SemiAnalysis openly urged AMD to circumvent Meta's internal infrastructure group and instead work directly with Meta's "TBD Lab" super-intelligence unit, which would presumably opt for the full MI455X .
• Micromanagement and repeated missteps: They pointed to a pattern of "catastrophic" infrastructure decisions, including a separate $2.5 billion acquisition that "ended in utter chaos" .

The core complaint is that by stripping down the chip, Meta sacrifices versatility and future-proofing for narrow near-term cost optimization, potentially locking itself out of more capable hardware needed for frontier AI workloads .

Not all observers agree. Some defense of the approach notes that recommendation systems do not need frontier LLM capacity. What they need is bandwidth per dollar on a task run billions of times a day, making the cut-down chip a rational specialization rather than a mistake .

Fit Within Meta's Five-Year, $60 Billion AMD Partnership

On February 24, 2026, Meta and AMD announced a five-year, $60 billion strategic partnership to deploy up to 6 gigawatts of AMD Instinct GPUs for Meta's AI infrastructure . Key terms include:

• Scope: AMD will supply custom MI450-based accelerators (the cut-down variant above), plus two generations of Epyc CPUs, across Meta's global data centers .
• Performance-based warrant: AMD issued Meta a warrant for up to 160 million shares of AMD common stock, at an exercise price of $0.01 per share. This represents approximately 10% of AMD's outstanding shares .
• Vesting schedule: The warrant vests in tranches tied to deployment milestones. The first tranche vests upon the initial 1 GW of shipments, with additional tranches unlocking as Meta scales purchases toward the full 6 GW .
• Price hurdles: Further vesting is tied to AMD's stock achieving price thresholds up to $600 per share .

The custom "Half-MI450" is the first concrete product manifestation of this megadeal: a chip tailored specifically to Meta's recommendation inference needs, designed to maximize cost- and power-efficiency per inference at massive rack-level scale, exactly what the $60 billion partnership was structured to enable .