Semiconductors & Advanced Manufacturing
SEMICONDUCTOR SIGNAL April 29, 2026
The AI infrastructure wave is going global — and power is the wall it's about to hit.
Today's signal reads like a map of ambition: data centers breaking ground in India, securing capacity in Israel, snapping up assets in Brazil, planning campuses in Norway and Italy. The infrastructure spending cycle that's been driving chip demand since 2023 is spreading fast. But underneath the expansion headlines, a more urgent conversation is forming: power is becoming the defining constraint, and the industry is scrambling to solve it in real time. One quiet story rounds it out — optical interconnects, the technology that moves data between chips using light instead of electricity, just pulled in serious venture capital.
The AI Build-Out Is Now a Global Construction Project
The geographic spread in today's announcements is striking. Google broke ground on a $15 billion data center project in India — one of the largest single infrastructure investments in Indian tech history. Anan Data Center signed a contract worth "hundreds of millions of dollars" to supply 40MW of AI infrastructure for Crusoe (a cloud company specializing in AI computing) in Israel. In Latin America, private equity firm I Squared Capital acquired Elea, a Brazilian data center operator, adding it to a portfolio that now spans four continents.
New greenfield projects are queuing up across Europe: Kitebrook — backed by a co-founder of the Carlyle Group — is targeting 500MW of capacity in Norway. KevlinX is eyeing Genoa, Italy. And in Texas, Core Scientific (a company that pivoted from Bitcoin mining to high-performance computing hosting) secured 300MW of capacity at a Pecos data center, with the first phase of HPC (high-performance computing — the kind of processing AI model training requires) hosting due live next year.
Why it matters for chips: Every one of these facilities will be packed with GPUs (graphics processing units — the chips that do the heavy lifting in AI workloads) and the networking silicon to connect them. The geographic diversification of data center spending also reflects a broader push to distribute AI compute beyond US borders, which has direct implications for where advanced chip supply chains need to reach — and for the geopolitics of who controls that infrastructure.
Power Is the Wall — And Everyone Knows It
Beneath the expansion headlines, a more urgent conversation is happening about energy. Several of today's items converge on the same problem: building data centers is now easier than powering them.
Core Scientific's announcement explicitly mentions a push toward "behind-the-meter" power — meaning generating electricity on-site rather than drawing from the public grid, a significant operational shift that signals just how constrained grid access has become. A separate item asks whether we're having a "hydro moment" in data centers, as developers eye hydroelectric power as a stable, high-capacity energy source. Norway, for example, generates nearly all of its electricity from hydro — which may explain exactly why Kitebrook is targeting it.
Two sponsored pieces — typically a leading indicator of where industry anxiety is concentrated — are both about grid management. EPC Power's "M System" uses technology called Agile Grid Forming to stabilize the violent power swings that AI workloads create (AI chips can jump from near-idle to full draw in milliseconds, which is genuinely hard on grid infrastructure). A second piece argues that data centers "must become grid interactive" — able to modulate their own demand in real time, acting as partners to the grid rather than passive consumers.
Then there's Utah. A planning commission delayed a decision on a project backed by investor Kevin O'Leary that is targeting 9 gigawatts of eventual capacity. For scale: 9GW is roughly the sustained output of 9 large nuclear power plants, from a single project, in a single state. It has already secured backing from Utah's Military Installation Development Authority.
Why it matters for chips: Power constraints are quietly reshaping which chip architectures win. Performance-per-watt has become a genuine business metric, not just a spec sheet number — when electricity is scarce or expensive, an efficient chip has a direct cost advantage. This is part of why AMD, and the custom silicon teams at Google and Amazon, are competing hard on efficiency rather than just raw speed.
Light Is the New Wire — Silicon Photonics Gets Serious Funding
The semiconductor-closest story of the day: OpenLight, a startup that designs custom photonic chips — chips that use light instead of electrical signals to move data — raised an additional $50 million in an extended Series A funding round.
Silicon photonics (integrating optical components onto standard silicon chips) is increasingly seen as a necessary technology for AI-scale data centers. The problem it solves: as you pack more and more GPUs into a single cluster, the copper wires connecting them become a bottleneck — they consume power, generate heat, and cap out on how much data they can carry over distance. Light-based interconnects can carry more data with less energy over greater distances.
The "application-specific photonic chips" OpenLight designs go a step further than general-purpose optical components — they're built for particular workloads, which typically means better performance and lower cost at scale.
Why it matters: If silicon photonics reaches scale, it changes both chip packaging and data center networking in fundamental ways. Intel, Cisco, and a cluster of well-funded startups are racing to own this transition. A $50M raise in an extended Series A signals that investors see it as early-innings — not yet winner-takes-all, but the consolidation phase is approaching.
Infrastructure Capital Is Maturing — and Eyeing the Exit
The data center sector is entering a phase of financial maturation. Csquare — the successor company to Cyxtera, a data center operator that went through bankruptcy — filed an S-1 with the SEC, signaling a planned IPO (an initial public offering, when a private company sells shares to the public for the first time). Latitude.sh, a Brazil-based cloud infrastructure provider, signed a $25.1 million, three-year AI capacity agreement with an unnamed US technology company.
The pattern: private equity has spent years buying, building, and consolidating data center assets globally. Now some of those bets are being positioned for public market exits. Whether public investors share private equity's enthusiasm for infrastructure-as-AI-play will be a real test of how deep the conviction in this buildout actually runs.
The Trend to Watch
The buildout is real and the capital is committed — but power is becoming the sorting mechanism that separates credible projects from wishful renderings. Projects that solve the energy problem (behind-the-meter generation, grid-interactive design, hydro-rich geographies) will have a structural edge over those that don't. At the chip level, watch silicon photonics: it's the quiet enabling technology that may determine how large AI clusters can actually scale before physics says stop.
TL;DR - The AI data center buildout is going truly global: Google is spending $15B in India, Crusoe is in Israel, and private equity is snapping up infrastructure from Brazil to Norway — wherever this build goes, chip demand follows - Power is now the binding constraint: Behind-the-meter generation, hydro sourcing, and grid-interactive design are moving from nice-to-have to existential — underscored by a single Utah project targeting 9GW of capacity - Optical chips are stepping into the spotlight: OpenLight's $50M raise backs chips that use light instead of electricity to move data — the technology that may determine how big AI clusters can get before copper interconnects become the bottleneck - Infrastructure capital is looking for exits: With Csquare filing for an IPO and consolidation continuing across continents, the market is about to find out whether public investors believe in the AI buildout as much as private equity does
Compiled from 1 source · 20 items
- Data Center Dynamics (20)