Semiconductors & Advanced Manufacturing
The AI model arms race just hit a new tempo — one major coding model per week for three months straight — and this week's chip news is really three stories layered on top of each other: what the GPT-5.5 launch reveals about which chips are actually doing the work, why the semiconductor ecosystem is quietly diversifying beyond Nvidia, and why the next binding constraint on all of it may not be transistors at all. It may be megawatts.
When "Trained on Blackwell" Doesn't Mean What You Think
Dylan Patel at SemiAnalysis delivers the most clarifying data point of the week buried inside a broader model landscape breakdown: despite OpenAI and Nvidia jointly claiming GPT-5.5 was "trained" on a 100,000-unit GB200 NVL72 cluster — Nvidia's newest Blackwell-generation AI server system, in which 72 accelerator chips are linked together at extreme speeds into a single computational rack — Patel is blunt that this "training" is post-training only. The base model itself was still pre-trained on Hopper, Nvidia's previous-generation H100/H200 chips.
This distinction is load-bearing. Pre-training is where the foundational neural network is built — months of exposure to enormous datasets, the most compute-intensive phase of building any large AI model. Post-training (which includes reinforcement learning, the technique used to make a model behave helpfully rather than just predict text) is expensive but shorter and more targeted. The GB200 cluster handling post-training tells us OpenAI has access to Blackwell at scale. The fact that pre-training is still on Hopper tells us the full Blackwell ramp — TSMC and Nvidia's production pipeline of next-gen chips flowing to customers — is still very much in progress.
The model release environment surrounding this is extraordinary. Patel counts at least one major new coding-focused model checkpoint per week for the past 3 months: Qwen3.6-Plus, Kimi K2.6, Composer 2, Gemini 3.1 Pro, and more. Each represents a significant compute investment. GPT-5.5 itself prices at $5 per million input tokens and $30 per million output tokens (a "token" is roughly ¾ of a word — this is how AI API pricing works), making it 2x more expensive than its predecessor GPT-5.4. That premium reflects real infrastructure cost. More models, more training runs, more demand for the chips that run them — the demand signal for advanced AI accelerators isn't softening.
Intel Is Back, and Meta Is Buying ARM Chips at Scale
After years of painful share losses to AMD in the server market and near-irrelevance in AI training (where Nvidia has been essentially alone at the frontier), Intel reported Q1 2026 revenue of $13.6 billion, with data center growth as the primary driver. Shares jumped 20% — a remarkable reaction for a company that has spent the better part of five years disappointing Wall Street.
Intel sells two things relevant here: Xeon processors (general-purpose server CPUs — the central processing units that run the broader workload in any data center, distinct from GPUs which handle the parallel math of AI training) and Gaudi AI accelerators (purpose-built AI chips competing with Nvidia's H-series and B-series). The data center surge is lifting both. The global infrastructure buildout — every hyperscaler, every sovereign AI project, every startup racing to build compute — needs server chips beyond just GPUs, and Intel makes those.
Meanwhile, Meta — whose AI infrastructure spans training massive models like Llama, running billions of inference calls daily across Facebook and Instagram, and building its own custom silicon — signed a "multibillion-dollar" agreement with AWS to deploy "tens of millions" of Graviton5 cores. Graviton5 is Amazon's internally designed ARM-based server processor — a CPU-class chip (ARM is a chip architecture originally developed for mobile devices, now expanding rapidly into servers because of its power efficiency) built for high-throughput, efficient compute rather than the massively parallel math GPUs are optimized for. Meta is deploying this at enormous scale specifically for "agentic AI workloads" — AI that takes sequences of autonomous actions rather than just responding to single prompts.
The signal here is structural: not all AI compute is GPU compute. Inference — running a finished model to produce outputs — can often be handled more cost-effectively on efficient CPU-style chips, especially for workloads where memory bandwidth and throughput matter more than raw parallel horsepower. Meta's Graviton5 bet is a quiet but significant data point in the ongoing story of the chip ecosystem diversifying away from pure Nvidia dependence.
The Megawatt Problem: Power Is Now the Binding Constraint
The semiconductor industry has spent years consumed by where chips come from — TSMC's fabs in Taiwan, the US CHIPS Act trying to build domestic manufacturing capacity, China's efforts to develop its own advanced logic chips. The constraint is shifting. Now the question is where the power comes from.
Amazon-backed X-energy, a small modular reactor (SMR) company — SMRs are factory-built nuclear power plants designed to be deployed faster and cheaper than traditional nuclear facilities, which take 10+ years and tens of billions of dollars to construct — raised more than $1 billion in a public IPO, selling 44.3 million shares in an offering that was upsized due to demand. That a pre-revenue nuclear startup can command that kind of public market valuation tells you how acutely data center operators feel the power shortage.
The demand is real and immediate. Wärtsilä, a Finnish industrial company specializing in gas-powered generation equipment, signed two natural gas supply deals with data center operators in Ohio and Texas totaling more than 1 gigawatt of combined capacity — roughly equivalent to the output of a full-scale nuclear plant, contracted specifically to power AI infrastructure in two US states. That's a staggering number.
At the same time, the social license to build is fraying. A proposed $4 billion data center in Nobles County, Minnesota was rejected outright — county officials voted to block data centers from being added as a permitted land use category. A UK report this week flagged data center development as stymied by "inadequate community engagement." Resistance is cropping up in rural America and suburban Europe simultaneously.
This is the bottleneck that chip roadmaps don't account for. TSMC can build more fabs on a multi-year timeline. Nvidia can design more GPUs. But power infrastructure is local, politically contentious, and slow — transmission lines, substations, community approvals, grid interconnection queues that stretch years into the future. The semiconductor industry's capacity to produce chips is no longer the rate-limiting factor it once was. The rate-limiting factor is increasingly watts.
What to Watch
The model proliferation Patel documents will continue to drive demand for compute at every layer — training, post-training, inference. Intel's Q1 suggests the non-Nvidia server chip market is healthy, and Meta's Graviton5 deal is an early signal that the inference stack is diversifying in ways that could matter to Nvidia's long-term pricing power. But both of those stories are ultimately bounded by the same constraint: you cannot run chips you cannot power.
The real trend to track over the next 12–18 months is whether the power infrastructure gap — SMRs, grid expansion, gas generation, community approval processes — can keep pace with the chip supply curve. If it can't, you'll see AI compute capacity constrained not by fabs or chip design, but by something far older and more stubborn: the electrical grid.
TL;DR - GPT-5.5 was post-trained on Nvidia's newest Blackwell chips but pre-trained on last-gen Hopper — a subtle but important signal that the full next-generation chip ramp is still mid-flight, even as model releases are accelerating to weekly cadence - Intel posted its strongest data center quarter in years ($13.6B revenue, shares +20%), and Meta's massive deal to deploy "tens of millions" of AWS's ARM-based Graviton5 chips for AI workloads shows the compute ecosystem is quietly diversifying beyond pure Nvidia dependence - Power is replacing silicon as the binding constraint on AI infrastructure — a nuclear startup just raised $1B at IPO, data center operators are signing gigawatt-scale gas deals, and communities in Minnesota and the UK are starting to say no to the buildout entirely
Compiled from 2 sources · 21 items
- Data Center Dynamics (20)
- Dylan Patel (1)