Deep

OLIX raises $312M to challenge Nvidia's inference chip monopoly

HyperSinc Intelligence/AUGUST 4, 2026

UK startup OLIX closed a $312M Series B at $3.3B valuation with backing from Arm and Netflix co-founder Reed Hastings, betting photonic chips that eliminate costly memory can compete against Nvidia in AI inference.

OLIX raises $312M to challenge Nvidia's inference chip monopoly

4 min read

TL;DR

→OLIX raised $312M Series B at $3.3B valuation, tripling from $1B six months prior, with Arm as strategic investor.
→The bet: custom photonic inference chips that bypass expensive high-bandwidth memory, a direct structural challenge to Nvidia's data center economics.
→Watch: DX-1 chip delivery to first customers in H2 2027; whether Arm's participation signals real OEM adoption; whether photonic interconnect scales beyond 100K chips.

OLIX Computing just tripled its valuation in six months, from $1 billion in February to $3.3 billion on Monday, by raising $312 million in a Series B that landed backing from Arm, Netflix co-founder Reed Hastings, and a roster of tier-one venture funds. The headline number alone signals market confidence. But the real story is architectural: OLIX is building inference chips that eliminate high-bandwidth memory, the most expensive and supply-constrained component in Nvidia's data center stack. If that works, it does not just compete with Nvidia. It breaks one of the structural moats that has made Nvidia's inference business nearly untouchable.

The core innovation sits in OLIX's DX-1 chip, a decode accelerator designed for the inference stage where models generate output tokens in response to user input. Instead of offloading model weights to external HBM, the expensive, scarce memory that drives up cost and latency, DX-1 stores models directly in on-chip SRAM and uses optical interconnects to communicate across a massive array of chips. For 100-billion-parameter models, OLIX claims DX-1 can deliver over 10,000 tokens per second per user while consuming less power per token than general-purpose GPUs running large batch sizes. Those numbers are not hypothetical: they are the threshold at which inference becomes cheaper to run in aggregate than keeping GPUs busy with GPU memory bottlenecks.

Arm's participation in this round is the signal that matters most. Arm does not invest in point solutions, it invests in architectural shifts that reshape how silicon gets designed and deployed. Arm's intellectual property underlies the overwhelming majority of processors manufactured globally. If Arm sees OLIX's approach as a viable path forward for inference, it sends a message to every OEM and cloud provider that the assumption, inference defaults to Nvidia for the next five years, is no longer safe. The company has also appointed Nick McKeown, the Stanford computer scientist and co-inventor of software-defined networking and OpenFlow, to its board. McKeown's credential is not ornamental: it signals that OLIX is not just optimizing existing chip design patterns but reimagining how compute distributes across large interconnected systems. That is a fundamentally different engineering problem than building a faster GPU.

The funding timeline is aggressive but credible. OLIX says it will deliver DX-1 to first customers by the second half of 2027, roughly 14 months from now. That is fast enough to matter for 2027 data center procurement cycles but slow enough to suggest the company has actual tape-out roadmaps, not just simulation results. The company is also hiring across six cities, London, Bristol, Austin, Toronto, San Francisco, which indicates serious intent to support customers and manufacturing partners, not just evangelize an architecture.

Here is what does not happen if OLIX executes: Nvidia's inference business does not collapse. But the margin structure changes. If photonic inference chips can deliver OLIX's claimed efficiency gains at scale, cloud providers face a real economic choice, not just a default procurement path. That choice itself, having optionality, breaks the dynamic that has sustained Nvidia's pricing power in inference. The semiconductor index is down 20 percent in the past month on concerns about AI spending slowdowns, but that anxiety has not infected private capital. OLIX's round proves the opposite: the best capitalists in the world are still betting $312 million that the inference layer is not settled, and that the settlement requires photonic silicon, not faster conventional chips.

Watch three things: whether DX-1 ships to actual customers on schedule in H2 2027 and achieves the claimed token throughput in production workloads; whether Arm announces OEM partners publicly that commit to designing systems around OLIX's architecture; and whether the claimed scaling to 100,000 interconnected chips actually reduces latency variance in real data center conditions. If all three happen, Nvidia's inference business enters a genuinely competitive era. If any one slips, the photonic thesis remains a high-conviction bet with an execution problem.

Key Takeaways

OLIX is not pitching incremental GPU improvements, it is attacking the supply constraint that makes Nvidia's inference business defensible: high-bandwidth memory costs and allocation.
Arm's participation as a new strategic investor signals that the IP backbone of global chip design sees OLIX's architecture as a legitimate threat to the current inference hierarchy, not a niche alternative.
If OLIX delivers DX-1 on schedule and achieves the claimed 10,000+ tokens per second per user at scale, it fractures the assumption that inference economics default to Nvidia for the next five years.

What it meansA photonic inference alternative backed by Arm and $312M in capital could reshape data center procurement away from GPU-centric architectures within 18 months of customer delivery, directly pressuring Nvidia's inference-chip pricing and availability leverage.

DISCLAIMER

This article is for informational purposes only and does not constitute financial, investment, legal, or tax advice.

SOURCES

← Back to HyperSinc