Chips Crossed $1 Trillion. Now the Industry's Biggest Problem Is the Power Grid.

On Tuesday morning, the semiconductor industry threw itself a champagne toast at the Moscone Center in San Francisco. The occasion was historic: global chip revenue has crossed $1 trillion a year for the first time, according to the Semiconductor Industry Association. It was also a farewell. SEMICON West, the industry's flagship trade show, was opening in San Francisco for the last time after more than five decades there. Next year, the whole show moves to Phoenix.

The mood was celebratory. The subtext was not. Read the program for this year's show, running October 13 through 15 under the theme Transform Tomorrow, and the transformation on everyone's mind is not smaller transistors. It is electricity. The industry toasted its biggest revenue milestone ever while its own agenda asked, in increasingly direct language, how AI's power appetite gets fed.

$1 trillion, four years early

The number arrived ahead of schedule. As recently as 2022, forecasters expected the chip industry would not clear $1 trillion in annual revenue until 2030. It got there in 2026. The SIA reports that 2025 industry sales reached $791.7 billion, and its forecast of 26% growth in 2026 carries the industry past the threshold.

The engine is AI infrastructure. Hyperscalers have been ordering AI-optimized servers by the tens of thousands, and each carries dramatically more silicon and draws dramatically more power than the enterprise equipment it replaces. Logic chip revenue alone rose 40% in 2025, reaching $301.9 billion, as GPU and accelerator demand reshaped the product mix. Memory is the other half: Gartner forecasts DRAM revenue will rise by more than 200% in 2026, driven by the high-bandwidth memory stacked onto AI accelerators.

SEMI Americas President Joe Stockunas called the $1 trillion clearance evidence of remarkable demand, while cautioning that growth at this scale presents new challenges the industry must tackle together. That duality set the tone: more than 35,000 attendees and 734 exhibitors across a sold-out floor of 1,500 booths, gathered to celebrate a milestone nobody expected this decade, and to confront the constraint that could slow the next one.

The industry toasted $1 trillion in San Francisco this week. Its own program agenda says the next trillion will be decided by the power grid.

Goodbye, San Francisco

SEMICON West has been a Bay Area fixture for more than 50 consecutive years. Starting in 2027, it relocates permanently to the Phoenix Convention Center, with the first edition scheduled for March 30 through April 1, 2027, on a new spring calendar. SEMI Americas formalized the permanent move in July 2026.

The reason is physical, not sentimental: the fabs moved. TSMC's North Phoenix campus is now the most advanced semiconductor manufacturing complex on US soil, with Fab 21 in high-volume 4nm production and a total announced investment of $265 billion covering plans for ten wafer fabs, two advanced packaging facilities, and a research center. The 2025 Phoenix edition drew 64% more attendees and 45% more exhibitors than the 2024 San Francisco edition, a data point that made the move easy to sell.

The federal government has come along for the ride. This year's CEO Summit keynotes share a stage with senior officials from the Commerce Department, the State Department, and the National Science Foundation, a measure of how thoroughly the CHIPS Act has turned chip manufacturing into national policy. The grant ledger: $6.6 billion to TSMC's Arizona buildout, $8.5 billion to Intel, and hundreds of millions more to Amkor, Micron, Samsung, and others.

The energy wall

The most telling session on the agenda runs Wednesday afternoon: "Future of AI Computing: Bending the Curve for Radical Efficiency Improvement." The framing from SEMI is explicit. Sustaining AI infrastructure growth is now an energy and power problem, and solving it requires collaboration across the entire supply chain rather than isolated advances in any single segment.

The numbers explain the urgency. The International Energy Agency projects that global data center electricity consumption will reach roughly 1,000 terawatt-hours in 2026, about the same as the entire country of Japan consumes, and that the figure will double by 2030, with AI-focused power use tripling over the same period. One AMD executive put the constraint in investment terms at SEMICON China earlier this year: each gigawatt of power corresponds to roughly $25 billion in data center facility investment, which makes the grid, not silicon capability, the binding limit on AI's expansion. Some analysts project that 30 to 50 percent of planned data center capacity for 2026 will slip into 2028 specifically because of grid interconnection delays and power availability shortfalls.

The industry's technical answer runs through advanced packaging. Across all three days, the Advanced Packaging and Heterogeneous Integration Summit centers on co-packaged optics, the architecture the industry is betting on to keep data moving without burning ever more power. The logic is straightforward: as data rates push past 800 gigabits per second per lane, routing signals through copper traces to optical modules at the edge of a chassis degrades signal integrity and dissipates power. Co-packaged optics places the optical transceiver directly inside or adjacent to the chip package, so the switch ASIC or GPU communicates optically over very short distances instead of electrically over the full path. Less distance, less latency, less wasted power.

The market math is striking. IDTechEx projects the co-packaged optics market will exceed $20 billion by 2036, growing at 37% a year, and NVIDIA already commercialized the approach with its Spectrum X Photonics and Quantum X Photonics switch platforms at GTC 2025. Meanwhile, a Day 1 session titled "Thermal is the New Node" captures the reordered priorities in five words: managing heat inside a stack of heterogeneously integrated chips is now the engineering challenge that determines performance, not transistor geometry.

What the road to $2 trillion actually requires

The rest of Wednesday's program reads like a checklist of everything the $1 trillion party glossed over. A supply chain resilience session examines visibility across geographies and manufacturing tiers, with AI, digital twins, and network planning pitched as the tools for reducing risk while sustaining the buildout pace. A Bulls and Bears panel brings Wall Street into the room, with analysts from Bernstein Research, Cantor Fitzgerald, and others debating whether the AI-driven capital expenditure cycle is sustainable, a question markets asked loudly after the October selloff that dragged the Philadelphia Semiconductor Index down roughly 2.6% on the 12th.

A dedicated Cybersecurity Forum on Wednesday addresses the fab and its supply chain as a target for state-sponsored and criminal cyber actors, because a successful attack on the process control systems of a $10 to $30 billion facility running around the clock under remote software control means physical production disruption, not just data loss.

Which brings the perspective into focus. The $1 trillion milestone was about demand, and demand came from AI. The $2 trillion target that SEMI has now named will be about something else: physics, power, and people. The industry spent fifty years learning how to print ever smaller transistors. The defining skill of the next decade may be negotiating with the electric grid.