Memory makers sell out to 2029, a photonic switch for AI racks, and 5 new chip ideas

Optical circuit switches cut power and latency in AI clusters by skipping optical-to-electrical conversion. Google already runs MEMS-based switches. PsiQuantum offers a chip-scale silicon-photonics version from a standard foundry. The results are company-reported lab results, not yet independent.

Your AI data centre has a traffic problem. Every switch turns light into electricity and back. PsiQuantum just showed a chip that switches the light itself, in 3 milliseconds. Samples are out. Sources: PsiQuantum, IEEE ComSoc.

HBM4 moves the base die to a logic process, which opens room for compute. DuoPD splits LLM prefill (GPU) from decode (near-memory). Gains are simulation-only and need memory-vendor buy-in.

AI answers are slow because data shuttles back and forth from memory. A Seoul team put tiny processors inside the memory stack. Simulated result: up to 2.2× faster answers. Real chips needed. Source: arXiv.

Multiferroics could let voltage, not current, flip magnetic bits, cutting memory energy. This paper shows the effect survives at the scale real devices need. It is fundamental physics; manufacturing is far off.

Chips waste energy pushing current to flip bits. A material that flips magnetism with voltage could fix that, but it died when made thin. Brown just made it work at 1.6 nanometres. Source: Nature Materials.

FPGAs fill gaps where AI networking and pre-processing change too fast for custom chips. The figure is vendor-claimed. Altera is private, so there is no direct listed stock.

Some chips can be re-wired after they ship. Altera's new Agilex 2X claims 1.8× more work per watt for AI networks. Source: Altera.

Licensable NPUs let many companies add AI without designing an engine from scratch. Gains are vendor-claimed and revenue arrives through royalties years later.

Not every company can design an AI chip. Ceva sells the design. Its new NeuPro claims double the work per area. Ships to licensees in 2027. Source: Ceva.

Optical AI compute is still pre-revenue but backed by granted IP and an MIT software partnership. Watch for tape-out news; foundries with photonics lines are the picks-and-shovels play.

What if your AI chip ran on light? A US startup called Neurophos is doing AI math by shining light through microscopic patterns. Less heat, more speed. It's early: no product yet. But they hold patents and just teamed up with MIT. If it works, photonics foundries win first. Source links on our page.

Early-stage photonic interposer aimed at the 'memory wall'. High upside, high execution risk; laser and photonics suppliers benefit regardless of which startup wins.

AI chips are starving, waiting for data from memory. Volantis wants to fix that with light: an optical highway built right under the chip. It's just a plan for now. But if it works, laser makers like Coherent and Lumentum get a new market. Links on our page.

IBM is pairing research with patent filings, signalling commercial intent. Long-dated; relevant to memory makers and analog-AI specialists.

Imagine one memory cell holding several numbers at once. IBM just published that trick, and filed a patent on it the same season. That's a signal they plan to use it. It's years away, but memory makers should pay attention. Sources on our page.

A memory giant publishing fixes for analog AI errors suggests in-memory compute is moving toward products. Micron is the direct beneficiary.

Analog AI chips are super efficient but sloppy. Micron found a simple fix: spread each number across several cells, like averaging friends' guesses. When a memory giant solves accuracy, products tend to follow. Link on our page.

Next-gen per-lane speed is a key spec for data-centre optics. Lab result today, but it maps directly onto transceiver vendors' roadmaps.

Data centres talk with light. A new chip-sized switch flips light 200 billion times a second on one colour. Faster switches mean cheaper, cooler AI clusters. Optics makers like Marvell and Coherent are the ones to watch. Link on our page.

A sober reality-check paper: 2D CFETs are a 2030s story. The tool makers win either way because every new transistor needs new deposition and etch.

Chips are running out of room, so engineers want to stack transistors made of material just one atom thick. Europe's top chip lab, imec, tested it, and says: not yet. That honesty is useful. The safe winners? The toolmakers. Link on our page.

Hybrid bonding is at the orders stage. Timing slipped for HBM4E, which favours legacy bonder makers short-term and Besi/AMAT longer-term. Adeia is an under-the-radar IP play.

Next-gen AI chips are glued together atom-close using copper. Toolmakers Applied Materials and Besi are teaming up to sell the machines. But memory giant SK hynix says it's not ready yet. Delay now, big wave later. Links on our page.

Glass substrates are moving from lab to pilot lines. Substrate makers and glass specialists are positioned; India gets a foothold via 3DGS's Odisha project.

The board under an AI chip may soon be made of glass. It stays flat and fits more wires. TSMC is testing it with Japan's Ibiden; Samsung and BOE are racing too. Even India has a glass-packaging plant coming. Links on our page.

On-device learning for edge AI is a real gap. FPGA demo today; small-cap neuromorphic and FPGA vendors are the exposure.

Most AI learns in giant data centres. This brain-like chip learns on the spot, and it beat Nvidia's Jetson by up to 3 times in tests. It's an FPGA demo, not a product. Watch small players like BrainChip. Link on our page.

Pure research; tracks a promising direction (reconfigurable, back-end-compatible devices). No investable pure play yet.

What if one tiny transistor could switch jobs: logic one moment, brain cell the next? Singapore researchers built exactly that from a material one atom thick. Decade-away science, but it's how the future starts. Link on our page.

Eye-catching lab gains; geopolitics may keep this inside China. Monitor for replication.

A Chinese lab says its light-based brain chip is over a thousand times faster on vision tasks. Big claim, lab test. If it holds up, photonics becomes the brain of edge AI. Link on our page.

The proof point that photonics research becomes revenue: ST is at volume. This is the near-term leg of the optics story.

Remember those light-based chip papers? Here's one that made it: ST now mass-produces silicon photonics for data centres. Paper to profit, in action. Link on our page.