Semiconductor AI Chips: HBM's Reign Is Ending

Semiconductor AI chips have a memory problem, and the market keeps pricing the wrong constraint. This is a documented trajectory, and the numbers read clearly.

London startup Olix just closed a $312 million Series B at a $3.3 billion valuation, the largest semiconductor venture round ever raised by a European company, per Tech Times reporting from August 2026. The thesis behind that capital is sharp.

AI inference stalls on memory bandwidth, and shuttling data with light beats pushing it through copper. The consensus keeps buying arithmetic throughput. That frame is wrong.

The consensus has the frame wrong

Every token a language model generates runs through two phases. Prefill processes the prompt in parallel and demands raw compute. Decode emits one token at a time and lives or dies on memory bandwidth.

That distinction is the whole game. In a standard GPU setup, model weights sit in high-bandwidth memory, and that memory must saturate on every decode step. Nvidia's H100 delivers roughly 3.35 terabytes per second, and decode drinks all of it.

Consider the arithmetic of a decode step. The chip fetches billions of weights, uses each once, and moves on. Compute sits idle while memory hauls data.

The market fixates on FLOPS. Marketing decks trumpet throughput while the real ceiling sits in the memory subsystem. The 90% of analysts tracking compute have the present right. They have the pace of change wrong.

The cost curve says who wins

Follow the physics of moving data. Copper interconnects burn energy and add latency as distance and bandwidth climb. Optical links flatten that penalty, moving bits as light at low energy per bit.

Olix's X-1 platform unrolls a model across specialized chips, assigning each chip to a single stage of the inference pipeline. A slow-and-wide optical die-to-die link stitches them together. The system scales to 10,000 chips in a rack-scale domain, delivered as a complete rack of chips, lasers, and optics.

This is a regime change, rather than a trend. When the interconnect stops being the constraint, the economics of inference invert. Bandwidth per dollar becomes the metric that matters.

Reed Hastings, Arm, and the UK Sovereign AI Fund joined this round for a reason. Strategic money follows the constraint, and the constraint moved from transistors to the wires between them.

Cliff event: inference stops being HBM's hostage

Adoption of new silicon architectures jumps, rather than creeps. The trigger is a cost crossover, and the crossover arrives when photonic interconnect delivers more usable bandwidth per dollar than HBM stacking.

I place that crossover inside the 2027 to 2028 window for high-volume inference deployments. The consensus pins it past 2030, treating photonics as a lab curiosity.

The gap between those dates is two-plus years, and the capital now flowing into the category closes it fast. When a $3.3 billion company ships complete racks, the integration risk that slowed adoption collapses.

Cliff event: photonic interconnect reaches inference cost parity, 2028, at hyperscale. Inevitable, and arriving sooner than the market models.

Three categories that change shape

Three parts of the stack will look different by 2029:

• HBM vendors: SK Hynix, Samsung, and Micron built enormous margin on stacked DRAM. Photonic architectures thin that dependency.
• GPU incumbents: Nvidia's moat rests partly on HBM supply and NVLink. Optical rack-scale challengers attack both flanks.
• Inference clouds: providers pricing tokens on GPU economics face rivals pricing on bandwidth per watt.

Each shift moves margin, and margin moves power. The vendor selling the scarcest resource captures the rent. For a decade that resource was compute. The next decade prices bandwidth.

Watch the memory makers closely. Their capital expenditure assumes HBM demand compounds forever. A photonic detour reprices those factories, and the reprice arrives faster than annual reports admit.

Application builders sit on the winning side. As the model layer commoditizes, cheaper inference expands the addressable market for every vertical product built on top.

My position, and what would change it

My position is direct. Photonic interconnect, rather than a further HBM generation, resolves the AI inference memory bottleneck, and value in the stack migrates from raw compute toward bandwidth efficiency and the application layer above it.

The reasoning follows the money and the physics. Capital chases the constraint that binds, and inference decode binds on memory. Light beats copper on energy per bit at scale, and that advantage compounds.

Here is the evidence that would break my thesis. Should HBM roadmaps deliver a step change in bandwidth per dollar that matches optical economics inside 18 months, the crossover slips. Should optical yields and laser reliability fail at rack scale, adoption stalls and the cliff moves out.

I weight those risks as real, yet secondary. The trajectory favors light.

The prediction

Prediction: at least one top-three hyperscaler commits a production inference deployment to photonic rack-scale interconnect within the next twelve months, and Olix ships X-1 racks to a named customer in that window.

Confidence: Medium-High. Horizon: 12 months. This is a market-timing call layered on a technology call, so I discount it accordingly.

Kill signal: Olix slips rack delivery past mid-2027, or a new HBM generation lands with a bandwidth-per-dollar gain that erases the optical advantage. Either event resets the timeline.

What this means for your desk

Read this through your own mandate:

• CTO / Chief Innovation Officer: audit inference stack assumptions built on HBM scarcity.
• Venture Capital: the impossible-looking bet is the interconnect layer, and the data supports it.
• Chief Strategy Officer: a three-year plan that assumes GPU-priced inference assumes a world that fades.
• Technology Procurement: a long GPU supply contract signed today may lock you to obsolete economics.

The common thread is timing. The present belongs to compute, and the near future belongs to bandwidth.

Position ahead of the crossover, rather than after it. The market prices this category as speculative today. That window closes as the first production racks ship. More theses sit in the blog index.