Semiconductors & Advanced Manufacturing
The invisible software that designs every chip on earth is quietly becoming one of tech's most durable businesses — and AI is turbocharging it.
Meanwhile, the data center construction boom has entered territory that's straining power grids, rattling regulators, and forcing AI companies into some eyebrow-raising infrastructure arrangements.
THE SOFTWARE LAYER NOBODY TALKS ABOUT
Before a single chip gets manufactured, it has to be designed — and that design work is almost entirely controlled by 3 companies most people have never heard of. Patel published a detailed market analysis this week on EDA, or Electronic Design Automation: the specialized software that translates a chip architect's intentions into instructions a semiconductor fab can actually manufacture. Think of it as the CAD software of the chip world, except orders of magnitude more complex and with essentially no substitutes.
The Big 3 — Synopsys, Cadence, and Siemens EDA — hold over 85% combined market share, and the business has posted positive revenue growth every single year for over a decade. In CY2025, Synopsys (which completed its acquisition of Ansys, the engineering simulation software company) generated $8B in revenue; Cadence came in at $5.30B; Siemens EDA at an estimated $2.2–2.5B. Combined, the Big 3 account for roughly $16B, with the broader EDA and semiconductor IP market totaling $18B when smaller vendors and emerging Chinese competitors are included.
What makes this remarkable is the growth rate. Patel's data shows EDA growing at a 13% compound annual growth rate (CAGR) — the rate at which a market compounds year over year — while overall semiconductor R&D spending grows at only 7%. That 6-point spread has been widening since 2018, driven by three forces: hyperscalers (the Googles, Amazons, and Microsofts of the world) designing their own AI chips in-house, the rising cost of verifying that designs actually work at advanced nodes, and the economics of emulation hardware (physical test systems that simulate chip behavior before fabrication begins).
EDA tools now consume 9–12% of total semiconductor R&D spending — rising to 12–15% when you include the IP licensing revenue that Synopsys and Cadence earn from selling pre-designed circuit building blocks that chip designers incorporate rather than build from scratch. Synopsys's IP business alone generated $1.7B; Cadence's exceeded $0.7B.
The moat here is structural and deep. These tools are embedded into the PDKs — Process Design Kits, essentially the rulebooks that foundries like TSMC publish describing exactly how to build chips on their manufacturing lines. When TSMC certifies a design tool for its 3nm process (3nm referring to the size of the transistors, where smaller means more processing power per chip), that tool becomes the de facto standard for anyone building on that node. Switching costs are enormous: Patel estimates that a design at 3nm costs Apple roughly $170–260M and Nvidia upward of $100M — you don't casually swap out your design software mid-project.
Synopsys's $35B acquisition of Ansys, whose simulation software models how chips behave under heat, stress, and electromagnetic interference, is a bet that the design and simulation markets are converging as chips get more complex. Cadence, which Patel describes as having engineered a remarkable turnaround to reach 44.6% operating margins, is pursuing a three-horizon strategy spanning EDA, system design, and AI-native tools. Part 3 of his EDA Primer, covering how AI is beginning to reshape chip design itself, is forthcoming.
31.7 GIGAWATTS AND COUNTING: THE DATA CENTER SUPERCYCLE MEETS REALITY
The numbers arriving this week from the data center world keep getting harder to process. Global data center capacity under construction hit 31.7GW in 2025 — more than double the prior year — according to a new industry report. For context, a gigawatt (GW) is roughly the output of a large nuclear power plant. The industry is, in effect, racing to build the equivalent of dozens of nuclear plants' worth of computing capacity simultaneously.
Dallas, Texas, took the top spot for the first time in the rankings — a reflection of the Sunbelt's land availability and the limits of traditional data center markets like Northern Virginia and Silicon Valley. Elsewhere, Singaporean firm Sembcorp submitted a planning application for a 280MW data center in Teesside, UK; Fleet broke ground on a 230MW development outside Reno, Nevada; and Google pledged $15B for infrastructure in New Florence, Missouri.
But regulators are pushing back. Oregon's energy regulator approved a new rate class specifically for large-load data centers, requiring them to bear the infrastructure costs they impose on the grid rather than socializing those costs across other ratepayers. It's a sign of a broader shift: localities and utilities that spent years competing to attract data centers are starting to ask who pays for the transmission lines, transformers, and grid upgrades these facilities demand. The DOE (US Department of Energy) launched a data center grid integration test-bed at the National Laboratory of the Rockies — a federally funded effort to figure out how facilities that consume enormous and variable amounts of power can be designed to stabilize the grid rather than stress it.
Nebius — a data center operator spun out of the Russian technology company Yandex — announced it would deploy Bloom Energy fuel cells to power its US facilities, with the first project covering up to 328MW. Fuel cells, which generate electricity through a chemical reaction rather than combustion, are increasingly attractive because they can be deployed faster than grid connections can be upgraded.
AI COMPANIES ARE PAYING EXTRAORDINARY SUMS TO SECURE COMPUTE
The SpaceX IPO filing this week contained a detail that stopped people in their tracks: Anthropic, the AI safety company backed by Google and Amazon, is set to pay Elon Musk's xAI $1.25B per month to rent data center space. That's $15B annually flowing from one AI company to another for access to the physical computing infrastructure needed to train and run large language models (LLMs) — the class of AI system behind ChatGPT and Claude.
Separately, Anthropic is reported to be in early talks with Microsoft about using Microsoft's custom Maia AI chips — purpose-built silicon that Microsoft designed in-house to reduce its dependence on Nvidia's GPUs (graphics processing units, which despite their name are now the dominant hardware for running AI workloads). The talks are described as early-stage and may not result in an agreement, but the signal is notable: frontier AI labs are actively exploring every option to diversify their chip supply beyond Nvidia.
OpenAI, meanwhile, launched a "Guaranteed Capacity" offering this week — essentially a long-term compute reservation system that lets enterprise customers lock in access to OpenAI's infrastructure in advance. In a market where compute availability can be a competitive bottleneck, selling forward commitments is a way to monetize scarcity and lock in large customers.
The picture that emerges is of AI companies in an expensive, urgent scramble for computing power — willing to pay figures that would have seemed implausible three years ago, and increasingly willing to source that compute from unconventional partners.
What to watch: The EDA market is the clearest example of a "picks and shovels" play in semiconductors — companies that profit from chip complexity regardless of which chip designers win or lose. As AI drives chip design costs and verification complexity higher, the 3 companies that own this software layer are structurally positioned to capture a growing share of an expanding R&D budget. Patel's forthcoming Part 3 on AI's impact on chip design itself may be the most important installment: if AI can automate portions of chip design, it could either expand EDA's market further or — longer term — begin to erode the human-labor-intensive portions that justify current EDA pricing.
TL;DR - 3 companies you've never heard of — Synopsys, Cadence, and Siemens EDA — control 85%+ of the software that designs every advanced chip on earth, and their business is growing faster than the chip industry itself because AI is making chip design exponentially more complex - The data center buildout is unprecedented in scale — 31.7GW under construction globally in 2025, more than double the prior year — but regulators are beginning to force the industry to pay for the grid infrastructure it requires - AI companies are paying extraordinary sums for compute access — Anthropic's reported $1.25B/month to xAI for data center space illustrates how acute the shortage of AI infrastructure remains, and why diversifying chip supply (including Anthropic's talks with Microsoft over custom chips) is a strategic priority
Compiled from 2 sources · 21 items
- Data Center Dynamics (20)
- Dylan Patel (1)