Jeffrey Shainline spent a decade at the National Institute of Standards and Technology tinkering with superconducting optoelectronic circuits—the kind of work that rarely makes headlines but occasionally rewrites the physics of computation. In 2024, he left to start Great Sky, a Boulder-based company with a proposition that sounds almost absurdly ambitious: fuse superconductors, photonics, and conventional semiconductors into a single platform to cut the energy cost of AI by orders of magnitude.
The timing wasn't accidental. Data centers in the United States consumed roughly 180 terawatt-hours of electricity in 2024—a figure that, according to BloombergNEF, could swell to 8.6% of the nation's total electricity supply by 2035. That kind of trajectory doesn't just raise eyebrows among environmentalists. It's forcing grid operators, chip designers, and founders alike to confront an uncomfortable reality: the current architecture underpinning artificial intelligence may not be sustainable at the scale the industry is racing toward.
Great Sky's bet is technical and specific. The company is building what it calls a monolithic platform that pairs Josephson devices—superconducting circuits that operate at cryogenic temperatures—with single-photon optical links. Shainline's earlier NIST research demonstrated synaptic events running at around 33 attojoules of dynamic energy per event, a fraction of what conventional CMOS manages. Of course, the devil lives in the details: those numbers don't yet account for the overhead of keeping everything cold enough to superconduct, or the practicalities of building a system that works outside a physics lab.
Still, the broader question Great Sky is posing—whether photonic computing can scale fast enough to matter—is one the entire industry is starting to ask.
When Physics Gets in the Way
AI infrastructure today is running headlong into fundamental limits, the kind that can't be solved with better software or clever optimization. The International Energy Agency projects that global data center electricity demand will reach approximately 2.6% of worldwide consumption by 2030 under baseline scenarios, with U.S. facilities shouldering a disproportionate share of that growth. Domestically, power demand is expected to nearly double from around 35 gigawatts in 2024 to 78 gigawatts by 2035—a surge that has state regulators scrambling. Georgia approved a major generation increase in early 2026, a signal of how quickly the grid conversation is shifting from abstract forecasting to immediate necessity.
The culprit isn't simply that we're doing more computation. It's the memory wall. Modern transformer models are memory-bandwidth limited during both training and inference, a constraint well-documented by analysts at TrendForce and others. High-bandwidth memory supply has been tight, with Meta's MTIA roadmap explicitly flagging HBM bandwidth as the bottleneck during the decode phase. The industry is pushing toward 1.6 terabytes per second and beyond with HBM4, but the stubborn reality remains: moving data between chips, across wafers, and between racks burns energy at a rate that makes raw compute look almost efficient by comparison.
This is the opening photonics has exploited. Optical interconnects promise higher bandwidth and lower energy per bit than copper, and the sector has moved with surprising aggression. Nvidia invested $4 billion into photonics suppliers Lumentum and Coherent in March 2026. Broadcom began shipping its Tomahawk 6 switch—capable of 102.4 terabits per second with native co-packaged optics—in 2025. Silicon photonics as a market segment is projected to grow from around $2.65 billion in 2025 to $9.65 billion by 2030, a compound annual growth rate of 29.5%, according to MarketsandMarkets.
Photonics for interconnect, in other words, is already happening. Photonics for compute—the more radical proposition—is still in the proving-out phase.
Three Pressures Converging
The shift toward optical solutions is being driven by three forces that are colliding more or less simultaneously: interconnect bottlenecks, energy efficiency mandates, and the physical limits of electronic fan-out.
Start with interconnect. As AI models scale, the cost of shuttling data between processing elements increasingly dominates total power consumption. Co-packaged optics attempt to address this by integrating photonic I/O directly onto silicon, eliminating the lossy electrical interfaces that add latency and burn watts. Lightmatter demonstrated 1.6 terabits per second per fiber in March 2026 and has partnered with Global Unichip on CPO design flows. Ayar Labs unveiled what it called the world's first UCIe optical chiplet in 2025, targeting scale-up architectures. LightCounting, a market research firm, notes that co-packaged optics and linear-drive pluggable optics are expected to double silicon photonics' market share from roughly 30% in 2025 to around 60% by 2030.
Energy efficiency is the second driver, and perhaps the more visceral one. A February 2026 arXiv preprint describing a photonic in-memory tensor core called SKYLIGHT claimed device-level energy efficiency of 23.7 TOPS per watt. The work is preliminary—system-level overhead remains unclear—but the appeal is obvious. Photonic circuits can, in theory, perform matrix operations with lower energy than their electronic counterparts, especially when co-located with memory to minimize data movement.
The third pressure is fan-out, a constraint that might sound esoteric but becomes critical at scale. Electronic interconnects hit physical limits in density and latency as systems grow larger. Single-photon communication, the approach Great Sky is pursuing, allows massive fan-out across multiple wafers using free-space or waveguide links. This design philosophy, rooted in Shainline's NIST work on superconducting optoelectronic networks, envisions neural networks operating at light speed with distributed state across a monolithic opto-electro-superconducting stack. It's elegant on paper; the question is whether it's buildable in practice.
The Contenders and Their Bets
Great Sky remains early. The company has raised approximately $5.4 million across multiple filings, the most recent dated September 26, 2025. It placed an order for cryogenic equipment from Danaher Cryogenics in September 2025—Shainline noted publicly that the gear would help "demonstrate the speed and scalability of our superconducting neural networks." The company has posted positions for neural algorithm developers and superconducting electronics experimentalists, signals of lab-scale build-out. But no public benchmarks, system-level demonstrations, or product timelines have been disclosed. This is, in other words, still very much a research play.
Great Sky's approach is arguably the most radical in the field. Competitors have taken more incremental paths, which may say something about the relative risks involved. Lightmatter, valued at $4.4 billion after a $400 million Series D in October 2024, focuses on photonic interconnects for existing AI chips. Its Passage platform is designed to accelerate rack-scale communication, and the company has aligned itself with industry standards like UALink for scale-up interconnects—a pragmatic choice that avoids the chicken-and-egg problem of building entirely new compute architectures.
Celestial AI has raised multiple rounds—$175 million in March 2024, with subsequent raises in 2025—to build what it calls a "Photonic Fabric," positioning optical links as both memory and compute fabric. Lightelligence demonstrated its PACE 2 optical compute accelerator with ecosystem support for ONNX, PyTorch, and TVM in March 2026, targeting inference workloads where the economic case for alternative architectures is clearest.
Snowcap Compute, which raised $23 million in June 2025, is also pursuing superconducting circuits, using Josephson junctions for digital acceleration. Like Great Sky, Snowcap faces the challenge of cryogenic infrastructure and helium supply—a risk that moved from theoretical to very real in March 2026 when Qatar's Ras Laffan helium complex shut down, temporarily removing roughly 30% of global supply. When your business model depends on keeping chips colder than the surface of Pluto, helium geopolitics start to matter.
The broader ecosystem, meanwhile, is maturing rapidly. Foundries are scaling silicon photonics capacity: UMC licensed imec's iSiPP300 platform with risk production targeted for the 2026-2027 timeframe, and TSMC verified its COUPE silicon photonics process in the first half of 2025. The Optical Internetworking Forum held sessions at OFC 2026 explicitly focused on optical interconnect specifications for AI, a sign that standardization efforts are gaining momentum.
What Comes Next
The near-term opportunity is relatively clear. Photonic interconnects are becoming table stakes for AI infrastructure, with the industry aligning around 1.6-terabit data rates and higher. Merchant silicon and co-packaged optics are shipping now, which creates a commercialization pathway for companies focused on optical I/O. This is the low-hanging fruit, and it's being picked.
The longer-term question—whether photonic or superconducting compute can deliver system-level efficiency gains that justify the integration complexity—is murkier. Great Sky's vision of fusing single-photon optics with superconducting neural circuits is grounded in solid physics. Shainline's NIST research demonstrated attojoule-scale device energy, numbers that look compelling in isolation. But the gap between device-level measurements and deployable systems is wide, and crossing it requires solving problems that aren't purely technical. Cryogenic operation adds cost and operational complexity. Helium supply remains structurally volatile. And as a 2025 perspective piece in Nature Communications Physics noted, lifecycle carbon modeling and full-stack co-optimization are still gaps in the photonics-for-AI narrative.
What should founders and investors watch? First, whether Great Sky and similar efforts can demonstrate integrated system benchmarks—not just device-level energy figures, but end-to-end TOPS per watt including cooling, control, and data movement. Device physics is one thing; systems engineering is another. Second, the trajectory of foundry support for hybrid integration. If TSMC or Samsung commit serious capacity to photonic-electronic co-packaging, that changes the calculus. Third, how quickly helium supply chains stabilize, or whether alternative cryogenic solutions emerge that don't depend on geopolitically fraught gas reserves.
The photonics wave is here, that much is clear. The question is how far it travels beyond the interconnect layer—and whether the most ambitious bets, like superconducting optoelectronics, can navigate the chasm from lab to production before the next generation of CMOS chips closes the efficiency gap. It's a race with no guaranteed winner, which is perhaps what makes it worth watching.
