Why this matters

Modern AI runs on a treadmill that is accelerating. Since 2020, the compute used to train frontier language models has doubled every 5.2 months — roughly five times faster than Moore's Law ever promised.^[1] The hardware underneath has improved, but nowhere near enough to keep up: leading machine-learning chips become about 40% more energy-efficient each year, while the compute demanded of them grows by 4 to 5x.^[2]

The gap is widening, and it's structural. Software efficiency gains pre-train better models per FLOP, but the underlying physics of CMOS — heat dissipation, interconnect bottlenecks, the end of Dennard scaling — sets a hard floor. You cannot make a transistor much smaller, much cooler, or much faster than today's are. The transistor was the workhorse of computing since 1947. It is approaching limits no amount of clever software can paper over.

Compute growth doesn't stay abstract. It shows up on the grid, in water tables, and on carbon balance sheets. Frontier training runs already consume tens to hundreds of megawatts of power — the load of a medium-sized power plant — and the power required to train them is doubling every year.^[3]

Aggregate the global trend and the picture sharpens. The International Energy Agency projects global data-center electricity demand will double from 415 TWh in 2024 to around 945 TWh by 2030 — slightly more than Japan's entire annual electricity consumption today.^[4] Electricity demand from AI-optimized data centers specifically is projected to quadruple over the same window.^[5] In the United States, data centers will consume more electricity by 2030 than the entire production of aluminum, steel, cement, and chemicals combined.^[4]

Water follows electricity. Training GPT-3 alone evaporated an estimated 700,000 liters of freshwater on-site at Microsoft's data centers, with millions more consumed off-site in electricity generation.^[6] Morgan Stanley projects AI-related data centers will consume over 1 trillion liters of water annually by 2028 — an eleven-fold increase from 2024.^[7]

The question is no longer whether we need new substrates. It's which ones can scale. Photonic computing — computing with light through silicon photonic integrated circuits — has emerged as the most compelling candidate for the linear-algebra core of AI workloads. Light doesn't dissipate the heat that electrons do when pushed through resistive wires; it can carry many wavelengths simultaneously; and silicon photonics is fabricated in the same CMOS-compatible foundries that already build the world's chips.

The numbers are starting to back the promise. Tsinghua University's Taichi chiplet, published in Science in 2024, demonstrated 160 TOPS/W — roughly an order of magnitude better than today's best GPUs at 20–30 TOPS/W.^[8] Recent reviews in Frontiers in Physics show photonic tensor cores reaching projected compute densities of 880 TOPS/mm², compared to single-digit TOPS/mm² for conventional CMOS accelerators.^[9]

Memristive devices fill the complementary gap — non-volatile in-memory computation that eliminates the data-movement energy tax that dominates modern AI inference. Hybrid systems combining photonic interconnect, memristive memory, and CMOS control are the design space LUCENT explores.