The short version: London's OLIX (formerly Flux Computing) just raised a $312M Series B at a $3.3B valuation — reportedly the biggest semiconductor VC round ever by a European company — to build optical inference chips that skip HBM entirely. Its first product, the DX-1, is a decode-stage chip using on-die SRAM instead of High-Bandwidth Memory, slated for H2 2027. You can't buy it for over a year, so the news isn't "switch hardware." It's a high-conviction bet that tells you where your inference cost is actually stuck — and what to measure now.

What happened#

Around August 3, 2026, OLIX closed a $312M Series B at a $3.3B valuation, roughly six months after a ~$220M round at a $1B valuation — more than tripling its valuation in half a year. Reported backers are an unusual mix: Arm (strategic), Hudson River Trading (quant), the UK government's Sovereign AI venture fund (state), and Reed Hastings as an angel.

The company builds photonic — light-based — inference chips it calls Optical Tensor Processing Units. The headline choice in its first product, the DX-1, is what it leaves out: no HBM. It uses on-die SRAM and targets the decode stage of LLM inference specifically.

Why "no HBM" is the whole story#

HBM — High-Bandwidth Memory — is one of the quiet reasons token prices don't fall faster. Here's the mechanism, because it's the part that matters for your bill.

LLM inference has two stages. Prefill ingests your prompt and is compute-bound: it processes all the input tokens in parallel and keeps the accelerator busy. Decode generates the answer one token at a time, and it's memory-bandwidth-bound — every single token requires streaming the model's weights and a growing KV cache back through memory. That sequential, memory-hungry pattern is exactly what HBM exists to feed, and HBM is expensive, supply-constrained, and stacked onto the priciest accelerators. We've written before about how the HBM squeeze pushes founders toward renting GPUs — OLIX is attacking the same bottleneck from the silicon side.

So a chip that can serve decode without HBM is a bet that the current cost floor is a memory-architecture problem, not a raw-compute problem. That's a real, specific thesis — and a $312M round is a lot of conviction behind it.

Is it real, or a slide deck?#

Treat it as credible but unproven. Photonic computing has decades of research and genuine physical advantages for moving data, but turning that into a manufacturable, programmable inference chip that beats mature GPUs on real models — with a software stack people can actually use — is the hard part, and nobody's done it at scale. A large round from strategic and state backers is a strong signal about direction, not a shipped product. Keep your near-term hardware decisions on things you can rent today; read the GPU rental price map and the CoreWeave vs Lambda vs Nebius comparison for what's actually available now.

What founders should actually do#

Nothing about 2027 silicon. The useful move is upstream and available today: measure whether your workload is prefill-heavy or decode-heavy. That single ratio decides whether any decode-optimized, HBM-free hardware would ever move your bill — and it already changes how you should size and rent GPUs, as we lay out in the prefill/decode GPU ratio. If your costs are dominated by generating long outputs, you're precisely the workload this class of chip is aimed at, and it's worth tracking. If your costs are dominated by ingesting long prompts, HBM-free decode hardware won't save you — and you should be optimizing context and caching instead.

Either way, the OLIX round is a reminder that the race to cut inference cost has moved past "buy more GPUs" and into "change the memory architecture." For the money side of that same equation today, see rent a GPU vs LLM API break-even.