Home โ€บ Blog โ€บ Beyond Silicon
AI Compute Stack ยท Part 18 ยท The horizon

Beyond Silicon: Quantum, Light, Brains in Chips and Orbit, and What Actually Comes After the GPU

Every earlier part of this series ran into a physical wall: power, memory bandwidth, interconnect, heat. Headlines promise escape routes: quantum computers, chips that compute with light, brain-like processors, energy-recycling logic and data centres in orbit. This final part looks at each one with the same question: which wall does it attack, what has actually been demonstrated, and when could it plausibly matter for AI?

FL
FrontierAGI Team
โš ๏ธ How this was researched This field is full of company roadmaps and modelled projections. We separate three kinds of evidence throughout: peer-reviewed results (for example Google's and Lightmatter's 2025 Nature papers), company claims and roadmaps (IBM's 2029 target, Extropic's efficiency models, NVIDIA's link specs) and analyst forecasts. Claims are labelled as such. The scorecard ratings and "when could it matter" dates are our editorial judgement. See Reading the numbers.
2029IBM's target for 200 error-corrected qubits (company roadmap)
200+ Tbpsoptical I/O per package, Lightmatter Passage L200 (2026)
1.15Bartificial neurons in Intel's Hala Point research system
4 TPUsin orbit on Google's first Suncatcher satellite (Oct 2026)

1. The walls, and who claims to break them

The candidates are not interchangeable. Each one attacks a specific limit, and some attack limits that are not the binding ones for AI. Click a wall to see which technologies target it.

Physical walls vs candidate technologies

2. The near future is still silicon (2026โ€“2030)

The surest gains this decade come from refining what already works. TSMC's A16 (1.6nm-class) process moves power wiring to the back of the chip ("backside power"), freeing the front for signals; volume production is due in late 2026, with products expected around 2027โ€“28. A14 follows in 2028, with a backside-power version (A14A) in 2029. Just as important is packaging: 3D stacking of logic on logic (TSMC's SoIC) and wafer-scale systems that join many dies and HBM stacks on one wafer (TSMC's SoW-X, slated for 2027; Cerebras already sells wafer-scale chips).

The silicon roadmap

Company roadmap dates as reported; product timing can slip.

The lesson from The Memory Wall and Interconnect Wars applies: most of the bottleneck is moving data, not doing arithmetic. That is why the most mature "beyond silicon" technology is the one that moves data: light.

3. Light: photonics

Photonics has two very different stories. Moving data with light is arriving now. NVIDIA announced Quantum-X and Spectrum-X Photonics switches with co-packaged optics (the laser-driven optical engine sits next to the switch chip instead of in a pluggable module) for 2026. Lightmatter's Passage L200 co-packaged optics, announced in March 2025 for 2026, comes in 32 and 64 Tbps versions and supports over 200 Tbps of I/O per package; in June 2026 Lightmatter joined NVIDIA's NVLink Fusion ecosystem. Analysts expect co-packaged optics in volume around late 2027 (a forecast, not a roadmap).

Computing with light is still research, but it got more credible in 2025. Lightmatter's Nature paper (April 2025) described a photonic processor that ran ResNet, BERT and an Atari reinforcement-learning agent at close to electronic precision, which earlier photonic chips could not reach. An MIT chip reported in Nature Photonics performed the key operations of a deep neural network optically, with over 92% accuracy on a classification task in under half a nanosecond. Neither is a product that can train a frontier model.

Why light helps data movement

Illustrative comparison of an electrical and an optical link as distance grows.

4. Quantum computing

Quantum computing has made real progress. Google's 105-qubit Willow chip produced "Quantum Echoes" (Nature, 22 October 2025): a physics measurement (an out-of-time-order correlator) using 65 qubits that Google says ran about 13,000 times faster than on the Frontier supercomputer, and that can be checked by repeating it, which Google calls a verifiable quantum advantage. IBM's roadmap targets Starling in 2029: 200 logical (error-corrected) qubits running circuits of 100 million gates, with a larger successor, Blue Jay, after 2033. NVIDIA's NVQLink (October 2025) connects quantum processors to GPU supercomputers for error correction and control; 17 quantum hardware makers and nine US national labs signed on, and NVIDIA quotes under 4 ยตs latency and 400 Gbps.

For AI specifically, the evidence is thin. There is no demonstrated end-to-end advantage for quantum machine learning on real data. Several proposed speedups were "dequantized": classical algorithms with similar data access matched them. And loading classical data into a quantum state is itself expensive, often costing more than the quantum algorithm saves. The realistic AI-adjacent uses are simulating molecules and materials (which can produce training data for AI) and AI helping quantum computers, for example decoding errors.

Is quantum useful for this AI workload?

Our assessment based on current literature. Click a workload.

The logical-qubit ladder

Physical qubits are noisy; useful algorithms need many error-corrected "logical" qubits. Roadmap targets are company claims.

5. Brain-inspired and analog chips

The brain runs on roughly 20 watts partly because memory and computation sit in the same place. Two families of chips copy that idea. Neuromorphic chips use spiking "neurons" that only consume energy when they fire. Intel's Hala Point, the largest such system, has about 1.15 billion neurons on more than 140,000 neuromorphic cores and draws up to about 2.6 kW. Analog in-memory chips store weights as physical properties of memory cells (for example the resistance of phase-change memory) and do multiply-accumulate in place; IBM has built research chips with tens of millions of phase-change devices. IBM's NorthPole takes a digital route, placing memory right next to compute.

None of these is a production AI accelerator yet. Efficiency claims are large but come from narrow benchmarks (image classification, small models) and vary widely by source. The open question is whether they can run transformer-scale models with the precision and programmability GPUs offer.

Why putting memory next to compute matters

Illustrative energy for one multiply-accumulate with its weight fetched from different places.

Relative values are illustrative orders of magnitude commonly cited in computer architecture: off-chip DRAM access costs far more energy than the arithmetic itself.

6. Thermodynamic and reversible computing

Two newer ideas try to cut energy at the level of physics. Thermodynamic computing (Extropic) uses circuits whose natural noise does useful work: "probabilistic bits" that flip randomly with controllable odds, suited to sampling from probability distributions, which generative models do constantly. Extropic tested its X0 chip in 2025 and lists Z1 for early access in 2026, with over 250,000 probabilistic bits on a small die drawing under a watt. Its claimed efficiency gains of several orders of magnitude over GPUs are modelled, not measured, and a published test so far covers a tiny eight-site model. The US Commerce Department issued a letter of intent for up to $75 million, which is not a final award.

Reversible computing (Vaire) avoids destroying information, because erasing bits costs energy. Its Ice River test chip, built on a 22nm process and tested in August 2025, changes voltages gradually and recovers charge through a resonator, "like a pendulum instead of a hammer". Reported results: about 30% less energy than a conventional chip on the same task, and energy recovery factors of 1.77 and 1.41 on test structures, figures that exclude overheads of a full system. Turning this into a useful processor is the hard part.

Hammer vs pendulum

Illustrative: conventional logic dumps charge as heat each switch; adiabatic logic swings charge back and forth and recovers part of it.

7. Orbit: an update

Our full analysis is in Data Centers in Space. Since then, the field moved from slides to hardware. Starcloud flew an NVIDIA H100 in November 2025 and ran Google's small Gemma model from orbit (claims that it "trained an LLM in space" are not clearly supported). Google's Project Suncatcher launched its first test satellite on 1 October 2026 on a SpaceX rideshare, carrying four TPUs and about 1 kW of solar power, to test radiation, launch vibration and cooling; a two-satellite mission with Planet is planned for 2027, and Google's paper sketches 81-satellite clusters a kilometre across. SpaceX filed with the FCC in January 2026 for up to a million satellites in an "Orbital Data Center System", and Starcloud has proposed up to 88,000.

The physics has not changed: launch cost, getting rid of heat in a vacuum, radiation and maintenance. Orbit attacks the power and land walls, not the memory or interconnect walls.

8. The scorecard

Each technology rated 1โ€“5 on how far along it is, how directly it helps AI workloads, and how strong the evidence is. Click a column heading to sort.

9. When could it matter?

Our estimate of when each technology could first carry a meaningful share of real AI work, not just a demo. Solid bars rest on shipped products or firm roadmaps; hatched bars rest on company claims or research.

Editorial judgement based on the evidence above. Not forecasts from the companies named.

Verdict The next decade of AI compute is still silicon: better processes, 3D packaging, wafer-scale systems and, above all, light for moving data. Those attack the real walls (memory and interconnect) and are already on roadmaps. Quantum computing is advancing fast but has no demonstrated advantage for mainstream AI; its likely contribution is scientific simulation. Neuromorphic, analog, thermodynamic and reversible chips are genuine research bets with large but unproven efficiency claims. Orbit is an answer to the power and land problems, at a cost the launch market has not yet met. Treat all of them as options on the 2030s, not plans for 2027.

๐Ÿ‘€ What to watch

๐Ÿ”ข Reading the numbers

๐Ÿ“š References