AMD's CPU Inference Bet, Cerebras Pressure & TSMC Capex Watch

Jamie Cole (0:00)

AI Hardware & Chips Daily News.

I'm Jamie Cole, thanks for joining me. Today, CPU Renaissance or AMD hype, assessing $200 billion TamKlang.

AMD's Lisa Su just made one of the boldest forecasts in semiconductor history. She's projecting the CPU total addressable market will exceed $200 billion by 2030, growing at 50% compound annual growth rate over the next four to five years.

That's a 50% upward revision to previous guidance. The driver, she says, is Agentic AI. The important distinction here is what Agentic AI actually requires. Unlike large-scale model training, which is overwhelmingly GPU-bound, Agentic workloads involve autonomous AI systems doing continuous reasoning, coordinating multiple models and making sequential decisions. Su's argument is that this favors CPU-dense architectures. AMD's Helion Agentic rack, shipping in November, puts 32,000-plus Zen 6 cores alongside 32 Mi350 GPUs in a single system. That's a meaningful CPU-to-GPU ratio shift. The real test is whether the market agrees. Inference now represents roughly 60% of AI compute demand. AMD has initial production systems deployed at OgenAI, Meta, Microsoft, Oracle and Anthropic. That's not a pilot list. The ramp runs through Q4 2026 into the first half of 2027 Here's the structural tension that matters. Nvidia's software ecosystem, CUDA, TensorRT and its inference-optimized GPU stack has spent years embedding itself into every major AI deployment pipeline. AMD's hardware case for CPUs and inference is coherent. The software moat question is not yet answered. If CUDA continues to dominate inference tooling, the 50% CPU-CARTHC thesis faces real friction regardless of what the hardware can do.

That said, the customer list is hard to dismiss. These aren't evaluation agreements. Production shipments to hyperscalers suggest the CPU inference thesis has at least partial validation in the highest stakes environments possible. Whether it scales to a $200 billion market by 2030 is a different, open question. Separate from the AMD story, but connected to the same AI infrastructure layer, is a warning that's repricing a different part of the market. Broadcom CEO flagged that capital access, not chip availability, may be the actual break on AI expansion. Intel echoed similar concerns. The immediate effect landed on optical suppliers. Applied up to electronics dropped 9 percent. Coherent and Lumentum each fell 7 percent. Here's the thing. The signal here isn't that AI demand collapsed. It's that the market is starting to price in scenario where funding constraints slow data center build out, independent of semiconductor supply. For Applied Optoelectronics specifically, the concentration risk is acute. Its top 10 customers represent 99 percent of revenue. In a constrained CAPEX environment, that dependency becomes a structural vulnerability, not just a statistic. Cerebras is trading at 217 times forward earnings. That's not a typo.

The company carries a $25.4 billion contracted revenue backlog. Core revenue hit $209.9 million, and cloud services surged 281 percent year-on-year to $126 million.

Full-year guidance was raised to $880 to $890 million, with management targeting triple that revenue by 2027

The stock dropped 5.1 percent anyway.

The pressure points are straightforward. A lock-up unlock puts 19.4 million additional shares into the market, and OpenAI, which anchors the customer base, is actively diversifying its chip supply. At 200-plus times earnings, there's no margin for customer concentration risk to materialize. TSMC earnings are due, and the watch items are specific. The question is whether $60 to $64 billion in capital expenditure guidance executes at the upper bound, and whether management provides any 2027 framework. CoWAS Advanced Packaging Capacity is the delivery proxy for AI chip shipments across Nvidia, AMD, and others. Expansion timing there directly determines when contracted AI silicon actually reaches hyperscaler racks. Consider this. Intel, meanwhile, expanded its manufacturing collaboration with Applied Materials covering 18A and 14A nodes, plus advanced packaging.

The partnership signals progress on Intel's Foundry ambitions. The unresolved parts is where the yield and cost on 18A materializes fast enough to compete credibly with TSMC. The GPUD series launch is tied to that Foundry push, so execution matters beyond just the fab roadmap. The two things worth tracking from here are straightforward. First, AMD's Q4 production ramp velocity. If Helion systems ship at volume across that hyperscaler customer list, the CPU inference thesis gets real evidence behind it. If ramp slows, the $200 billion TAM reverts to marketing math.

Second, watch where the Broadbom and Intel's capital constraint warnings show up in TSMC's Q4 capex commentary. If TSMC scales back or softens 2027 visibility, it's a signal that the funding constraint meritive has real teeth, not just sceptre noise.

The AI infrastructure layer is entering a phase where the supply constraint has shifted. Chips are available, capital allocation is the new variable. That changes which business models win, and it's worth keeping that frame close. Thanks for listening. This podcast was built using AI technology. A YesWe production.