AI's 2027 ceiling is set by the grid, not the chip. Optics buys back only a few percent.

The constraint on how much AI capacity comes online in 2027 is no longer how many accelerators can be made. It is how many megawatts can be connected to them. The evidence from the past three months points one way. Grid operators are short of capacity, the backstop meant to fix the shortfall has been delayed, turbines are sold out for years, and the efficiency levers people talk about most, including photonics, act on small slices of facility power. Chips are still tight in places. But the tighter limit, and the slower one to fix, is the grid.

Start with the largest US grid market. PJM's 2028/2029 capacity auction cleared at its $325/MW-day price cap and still came up 6,831 MW short of its reliability requirement. That followed a shortfall of about 6,500 MW the year before. PJM says these are the first auctions in its history in which the whole region fell short. The fix was a one-time Reliability Backstop Procurement for new supply. On September 29, FERC accepted it but suspended it for five months, effective no earlier than February 28, 2027, and ordered a paper hearing on parts it found may be unjust and unreasonable. The same order records that nearly 5,100 MW of the 5,250 MW load-growth forecast behind the previous shortfall came from data centers. A procurement that starts in 2027 does not deliver power plants in 2027.

Building your own power plant is not a quick bypass either. GE Vernova's Q3 2026 results show gas equipment backlog and slot reservations up from 116 GW to 122 GW, and the company expects at least 130 GW by year end. Slot reservations are not firm orders; 58 GW of the total sits in that category. A Reuters analysis published today reports that a new US gas turbine now takes six years to deliver on average, that worldwide orders exceed 100 GW against 60–70 GW of annual production, and that developers are buying used turbines and converted jet engines. Nuclear restarts face the same wires. Constellation's 835 MW Crane restart, contracted to Microsoft, was told by PJM that the transmission upgrades it needs could take until 2031. Constellation is seeking regulatory workarounds to keep a 2027 date. The IEA's June 2026 briefing counts 47 GW of on-site gas projects for US data centers in the pipeline. It names gas turbines, transformers, grid connection and new transmission lines as the slowest bottlenecks. Memory and advanced packaging, the chip-side limits, clear faster.

Meanwhile the money keeps arriving faster than the megawatts. In calendar Q3 2026, Alphabet spent $51.4 billion on property and equipment. Amazon spent $77.3 billion and still expects about $220 billion for the year. Microsoft reported $45.2 billion including finance leases. That is roughly $174 billion in one quarter from three companies, though each defines capex a little differently. The telling detail came from Microsoft's earnings call. CFO Amy Hood said capex came in lower because “some third-party capacity deliveries we expected in Q1 were delayed,” cut calendar-2026 spending to about $170 billion, and said “customer demand continues to exceed available supply.” Hood did not blame power. But when the cash is there and the GPUs are there, a delay in delivered capacity is usually a delay in buildings, power and grid hookups.

So where can efficiency come from? Begin with what optics can actually reach. Berkeley Lab's 2024 report to Congress estimates that network equipment held flat at about 3% of US data center electricity from 2014 to 2022. It projects 23 TWh by 2028, which is roughly 4–7% of the projected 325–580 TWh total. That likely undercounts optics, because the model assigns InfiniBand switches a flat 700 W and counts NVLink inside the servers. Nvidia's own framing is bigger. At GTC 2025, Jensen Huang said each GPU needs six pluggable transceivers drawing about 180 W, or 180 MW for a million GPUs, and claimed 3.5 times better power efficiency for co-packaged optics. Our rough arithmetic follows. Berkeley Lab's figures imply about 1 kW of average draw per H100-class GPU, or about 1.4 kW at the facility meter with 2023's average PUE of 1.4. Against that, 180 W is about 13%, and a 3.5 times cut saves about 130 W, or roughly 9%. The share is smaller for newer accelerators that draw more per chip. That is a real saving. It is not a new power plant.

The optics are now real hardware. In late July, TrendForce reported that Nvidia had begun shipping its Spectrum-X co-packaged-optics switch to select partners, and that Broadcom's version cuts power by up to 70% against pluggables. TrendForce does not expect the volume ramp until 2027–2028, held back by optical-engine yield and by competition with AI chips for the same advanced packaging. Nvidia does not report CPO unit volumes. Its supply bets are large: $2 billion each in Lumentum and Coherent in March, $4 billion in total. Some outlets report a total of $6.5 billion for Nvidia's photonics investments, but we found no primary source for that figure and do not use it. Our post-OCP piece covers the deployment questions in more detail.

The bigger facility-level levers sit elsewhere. Berkeley Lab projects average PUE falling from 1.4 in 2023 to 1.15–1.35 by 2028, mostly through liquid cooling and the move to hyperscale sites. Going from 1.35 to 1.15 frees about 15% of a site's power for compute. Nvidia's 800 VDC architecture, due with its Kyber racks in 2027, claims up to 5% better end-to-end efficiency and 45% less copper. These are vendor claims, not field measurements. Stack optimistic versions of all three and a fixed grid connection carries perhaps a quarter more compute. That is worth having. But demand is not growing by a quarter.

The largest gains still come from the chips and the software, through more work per joule from each process node, lower-precision arithmetic, and better serving. In a power-capped world, the metric that matters is tokens per joule at a fixed site, not peak FLOPS. That shifts inference engineering toward cutting the energy cost of each token: batching, caching, smaller models routed by difficulty, and fewer wasted reasoning steps. The IEA notes that its growth outlook rests on agentic workloads, which use far more energy per task than plain text generation. If agents grow faster than efficiency improves, the gains will be spent on more tokens rather than on less power. Interconnect matters most during decode, when data movement dominates. That is where photonics earns its place, as a per-token saving, not a grid fix.

Our forecast for 2027 is that the AI capacity that actually comes online will track energized megawatts, not chip shipments. Watch three things: what PJM files after the FERC order and whether a backstop auction runs in early 2027; the share of GE Vernova's slot reservations that turn into firm orders; and whether hyperscalers start reporting delays in powered capacity rather than in third-party deliveries. Optics will keep improving, and it should. It cannot add the megawatts the grid is missing.