In October 2025, Google Quantum AI published a result that should have stopped the presses: its Willow processor ran a verifiable physics computation, the Quantum Echoes algorithm, roughly 13,000 times faster than Frontier, the fastest classical supercomputer on Earth, and a second quantum machine reproduced it. The news cycle gave it about two days. A chatbot feature launch that same month got two weeks.
That asymmetry is the story. Since 2023, the generative AI boom has mutated into agentic systems and, this year, into what NVIDIA calls Physical AI: robots that reason and act in the world, with ABB, FANUC, KUKA, Figure and Medtronic all building on its stack. The public watches the software demos. States, labs and capital markets are fighting over something else. The substrate. Compute, light, qubits. Three supremacies, one arms race.
AI was never a software story
Start with the money. OpenAI's Stargate program targets $500 billion and 10 gigawatts of compute. A frontier gigawatt currently prices at $35 to 60 billion depending on whom you ask, Bernstein at the low end, Jensen Huang at the high one. Europe's entire €30 billion gigafactory tender, the continent's flagship sovereignty play with bids closing this November, buys roughly 0.6 to 1 gigawatt at those prices. One gigawatt. That is a down payment, not a strategy.
Then look at export controls, where the arms race becomes explicit. In January 2026, Washington moved H200-class chips from "presumption of denial" to case-by-case review, while imposing a 25 percent tariff on advanced AI chips not bound for the US supply chain. In April, the bipartisan MATCH Act proposed choking off chipmaking-equipment sales and servicing to SMIC, Huawei, CXMT and YMTC, with ASML's allied markets pressured into alignment. In May, the Commerce Department issued weekend guidance closing the subsidiary loophole: license requirements now follow Chinese headquarters wherever the entity sits.
And still it leaks. Chinese AI firms are renting advanced NVIDIA compute in Southeast Asian data centers, exploiting the gap between owning chips and accessing them, and lawmakers are now debating whether Washington should regulate remote cloud access itself. Meanwhile China holds its own with globally popular cheaper models while the chips are blocked at the border. The Council on Foreign Relations calls the current policy "strategically incoherent and unenforceable." I find it hard to disagree. (Read The New AI Chip Export Policy to China: Strategically Incoherent and Unenforceable by Chris McGuire).
When a government starts treating a commercial GPU like a munition, the technology argument is already over. The control boundary has shifted from hardware to compute access, and policy is chasing a moving target with a spreadsheet. My read: the restrictions slow China's frontier training runs at the margin, but they mostly accelerate two things Washington didn't intend, a domestic Chinese stack and a global gray market for compute.
Quantum: real, verified, strangely patient
Now the front nobody prices correctly. Willow demonstrated below-threshold surface-code error correction in late 2024, with an error suppression factor of 2.14 on a 101-qubit distance-7 code, still the most rigorous scalable QEC result published. USTC's Zuchongzhi 3.0 answered in 2025 with a sampling claim of 10^15 over the best supercomputer. Harvard and QuEra showed fault-tolerant neutral-atom computation with 448 qubits, and IBM's Nighthawk does real-time qLDPC decoding in under 480 nanoseconds.
Here is the honest position, and I say this as someone who wants quantum to matter: supremacy on contrived sampling tasks has been demonstrated repeatedly, and the gap on every single benchmark remains contested and provisional. Classical algorithms keep catching up. They always do, right up until they can't.
The roadmap that counts is error correction, not supremacy headlines. IBM published an end-to-end fault-tolerant framework targeting Starling by 2029: 200 logical qubits, 100 million gates. Its Kookaburra system, roughly 4,158 physical qubits on the modular System Two architecture, aims at quantum advantage on a useful workload by end of 2026. DARPA, which does not spend money on vibes, put $125 million into PsiQuantum under its Quantum Benchmarking Initiative, a program built around one beautifully bureaucratic question: can anyone build an industrially useful quantum computer by 2033?
If someone ask me whether to budget first for post-quantum cryptography migration or for quantum machine learning. Migration, I will tell. Not close. Quantum is a hedge against the 2030s, not a successor to the GPU in the 2020s. The slope of progress is genuinely encouraging. The intercept is still far away.
Photonics manufactures on existing lines
If you want a dark horse, look at light.
The binding constraint on AI scaling is not FLOPS. It is moving data, and paying the power bill for moving it. I run local models on my own hardware, and the wall I hit is memory bandwidth, never compute. Scale that observation by a million GPUs in a cluster and you understand why the smartest money in the Valley is buying photonics.
Lightmatter raised $400 million at a $4.4 billion valuation for Passage, a wafer-scale optical interposer that moves data between chips with light instead of copper. PsiQuantum went further: photonic qubits built on standard 300mm lines at GlobalFoundries, 99.7 percent fidelity sending quantum states across 250 meters of ordinary fiber, detectors running at 2 kelvin, a hundred times warmer than superconducting approaches require. The market noticed. One billion dollars at a $7 billion valuation in September 2025, with BlackRock, Temasek and NVIDIA's venture arm in the round, then $10.5 billion by May 2026, with utility-scale sites breaking ground in Brisbane and Chicago. Canada's Xanadu became the first pure-play photonics quantum company to list, via SPAC at $3.6 billion. France's Quandela raised roughly €70 million. The leader outraises the next two combined by more than 6x.
My position: photonics is the only contender that attacks the bottleneck that actually binds AI, interconnect and energy, while riding existing semiconductor manufacturing instead of demanding exotic new fabs. And notice the convergence hiding in plain sight. The same photonics that links GPUs in a rack also carries qubits across a quantum network. One substrate, two revolutions. If any supremacy sneaks up on the incumbents, it is this one.
Supremacy arrives as a utility bill
Which brings us to the eschatology. Ray Kurzweil, unmoved, reiterated in a June 2026 interview what he has said for a quarter century: human-level AGI by 2029, the Singularity by 2045. The industry has quietly reorganized around similar dates without using the word. Meta launched Superintelligence Labs in July 2026, co-led by Alexandr Wang and Nat Friedman, backed by a $14.3 billion stake in Scale AI, a shift from model competition to what one analysis called "AGI organizational capability" competition. NVIDIA, meanwhile, is building the off-ramp from software to atoms. GR00T N2, its next robot foundation model, tops the MolmoSpaces and RoboArena leaderboards and succeeds at new tasks in new environments more than twice as often as leading vision-language-action models. Its open reference humanoid, built with Unitree around a Jetson Thor module delivering 2,070 FP4 teraflops at as little as 40 watts, ships in October, with Stanford, ETH Zurich, Ai2 and UC San Diego already committed.
Step back and the sequence GenAI, agents, physical AI stops looking like a march toward a metaphysical event. It looks like vertical integration. Energy, silicon, interconnect, models, actuators: whoever owns the full stack owns the meter. Supremacy, when it comes, will not announce itself in a paper titled "AGI achieved." It will show up as infrastructure, metered and regulated, the way electricity did. The singularity framing is a distraction for people who don't read capex tables. AGI will be declared by whoever sends the invoices.
What to watch
Three tells for the next twenty-four months. Whether IBM's Kookaburra demonstrates advantage on a workload someone would actually pay for by end of 2026. Whether optical interconnect ships as standard equipment in the next rack-scale generation rather than as a premium option. And whether agentic and physical AI revenue starts covering the $35 to 60 billion per gigawatt the industry is pouring into the ground. If supremacy arrives as infrastructure rather than as an event, will anyone still call it the singularity?