Semiconductors & Advanced Manufacturing
The story underneath today's chip news isn't really about chips — it's about electricity. The AI buildout has created a power and cooling crisis so acute that energy infrastructure is now the binding constraint on the entire industry. Meanwhile, a materials science breakthrough out of Italy hints at one possible way out, and governments around the world are starting to realize their data center strategies are as much about national security as tech policy.
The AI Data Center Has a Power Problem — and Everyone Is Scrambling
The AI boom runs on electricity, and the industry is finally reckoning with how much of it. Several stories this week circle the same problem from different angles: AI data centers — sometimes called "AI factories," meaning purpose-built facilities that train and run large AI models rather than just storing files — are pushing power density beyond what traditional infrastructure can handle.
What does that mean in practice? A standard data center rack might draw 10–20 kilowatts of power. Modern GPU (graphics processing unit, the chips that power AI training) clusters can push 50–100+ kilowatts per rack, generating heat that air cooling simply can't dissipate fast enough. This week's coverage includes a primer on dielectric fluids — electrically non-conductive liquids that can be used to immerse server hardware directly, pulling heat away far more efficiently than air — and a sponsored deep-dive on why power and liquid cooling systems now need to be designed together from the ground up, not bolted on as an afterthought.
The energy supply side is equally stressed. Eos Energy is partnering with Turbine-X to combine zinc-based battery storage with natural gas generation specifically for US AI data centers — a sign that the grid alone can't be relied upon to deliver the consistent, massive power loads these facilities demand. Hitachi Energy and Samsung C&T are collaborating on high-voltage grid infrastructure (HVAC systems — here meaning High Voltage Alternating Current transmission, not air conditioning) to serve the same market. And Alphabet (Google's parent company) signed a new energy supply agreement with Indiana utility NiSource to power a large-scale data center in Northern Indiana, while also expanding an existing deal with Amazon — a reminder that the big tech hyperscalers are now effectively functioning as anchor tenants for regional power grids.
Anthropic — the AI safety company and maker of Claude — is building out a dedicated energy team, hiring Sana Ouji from Google to help "responsibly and rapidly scale an ambitious data center portfolio." The fact that a pure AI lab now needs a professional energy acquisition function tells you something about how capital-intensive this industry has become.
The Public Is Paying for AI Infrastructure, Often Without Knowing It
Two stories this week expose a less flattering side of the data center boom: the public costs are being systematically obscured.
In the US, a new report finds that more than a quarter of US states do not properly disclose revenue lost to data center tax abatements — meaning the property tax breaks and incentives that states offer to attract data centers aren't being reported in a way that lets citizens understand what they're giving up. GAAP (Generally Accepted Accounting Principles — the standard accounting rules that govern financial disclosure) requires this transparency; many states aren't following it.
In Europe, a lobbying effort by a coalition including Microsoft, Amazon, Google, and Meta successfully secured an EU provision keeping environmental impact data confidential. This matters because data centers are significant energy consumers — the European Commission separately reported 6.4 gigawatts of installed data center capacity across the EU, and that number is growing fast. If the environmental cost of that electricity consumption is shielded from public view, it's harder for regulators and citizens to weigh it against the benefits. The EU's Energy Efficiency Directive (EED) was supposed to create that visibility; the data exists in aggregate, but operators fought successfully to keep operator-level detail hidden.
Together, these two stories suggest a pattern: the data center industry has become expert at capturing public subsidy while minimizing public accountability.
Graphene Optical Chips: A Glimpse of What Comes After Silicon
The most genuinely semiconductor-forward story this week comes from Italy, where a startup has raised €211 million to develop graphene-based optical chips, with a pilot manufacturing facility targeted for 2028.
Unpack that: graphene is a form of carbon arranged in a single-atom-thick lattice, with electrical properties that silicon — the material in virtually every chip made today — can't match. Optical chips use light (photons) rather than electricity (electrons) to move data, which is dramatically faster and more energy-efficient for certain workloads, particularly high-bandwidth data transfer inside AI data centers. Combining graphene's superior electrical properties with photonic (light-based) data transmission is a research frontier that could eventually change how data moves between chips and between servers.
€211 million is serious money for a materials-science startup, and a 2028 pilot line suggests this is past the "interesting idea" stage. It won't disrupt the silicon foundry ecosystem anytime soon — TSMC (Taiwan Semiconductor Manufacturing Company, the world's dominant chip manufacturer) isn't worried yet — but it's worth tracking as a signal of where post-silicon materials R&D is concentrating.
Sovereignty Is Now a Data Center Pitch
The week's geographic spread is striking: Portugal approved a National Data Center Plan, Morocco is building "high-performance" data centers to monitor its general elections, Brazil's São Paulo got a new Equinix facility, and — most unusually — Swedish defense company Saab pitched the Canadian military on a bundle deal combining fighter jets and a sovereign data center, positioning local data infrastructure as a national security asset.
That Saab pitch is worth pausing on. The idea that a weapons manufacturer's competitive differentiator now includes data center sovereignty is a sign of how thoroughly governments have concluded that who hosts your data — and where — is a strategic question, not just a procurement one. Countries that once outsourced all of this to US hyperscalers are now writing national plans and attaching data infrastructure to defense relationships.
What to Watch
The power constraint is the story of the next several years. Every ambitious AI roadmap — more parameters, faster inference, bigger models — requires more electricity. The companies that solve the power and cooling problem (or lock up energy supply agreements early) will have a structural advantage. The graphene optical chip story is the long-horizon bet: if data movement inside AI clusters can be made dramatically cheaper and faster, it unlocks a different kind of scaling. Watch for 2028 as a milestone year on that front.
On the policy side, the collision between the data center industry's appetite for opacity and governments' growing interest in sovereignty and accountability is just beginning. Expect more mandatory disclosure rules, more national data center strategies, and more tension between hyperscaler lobbying and public interest.
TL;DR - AI data centers are hitting a hard physical limit on power and cooling — the whole industry is racing to solve it with new fluids, batteries, grid infrastructure, and energy deals - The public is subsidizing data center growth through tax breaks and bearing its environmental costs — and in many cases isn't being told the full story - An Italian startup raised €211M to build graphene optical chips, a materials bet on what computing looks like after silicon - Governments worldwide are treating data center infrastructure as a national security issue, not just a tech procurement question
Compiled from 1 source · 20 items
- Data Center Dynamics (20)