James Dacombe doesn't look like someone who should be running two deep-tech hardware companies. At 25, the London-based founder dropped out of school at 16—no university degree, no traditional credentials—and yet he's just closed a $220 million funding round that values his photonics chip startup, Olix Computing, at just over $1 billion.
His pitch? Sidestep the entire infrastructure choking the AI chip industry.
While Nvidia, AMD, and a parade of hyperscalers scramble for limited supplies of high-bandwidth memory and advanced packaging capacity, Olix is betting on a different physics entirely: photonic chips paired with SRAM that, Dacombe claims, can deliver better economics for AI inference than the GPU architectures currently dominating data centers. It's a wager that arrives at a curious moment—just as the industry's center of gravity lurches from training gargantuan models to deploying them in the wild, where every token generated costs real money and every millisecond of latency matters.
The February 2026 round, led by Hummingbird Ventures with participation from Plural, Vertex Ventures, LocalGlobe, and Entrepreneurs First, brings Olix's total funding to roughly $250 million. First customer shipments are targeted for 2027, assuming the company can deliver on technical promises that remain, for now, largely unproven outside its labs. Olix—which rebranded from Flux Computing in January—now employs over 70 people across London, Bristol, Austin, San Francisco, and Toronto, with plans to more than double headcount by year-end.
The audacity extends beyond the technical gamble. Dacombe is simultaneously running CoMind, a separate venture developing non-invasive laser-based cerebral monitoring technology that's raised over $100 million toward a planned 2027 commercial launch. Two deep-tech hardware companies. Two wildly different problem sets. One founder who can't yet rent a car in most U.S. states without extra fees.
"Most founders would struggle to execute on one," observes a venture partner familiar with both companies, speaking on background. "He's doing two. It's either genius or madness, and we won't know which for at least another eighteen months."
When Training Stops Being the Story
The AI chip market is undergoing what industry insiders call a "fundamental reorientation"—though that polite phrase obscures a more chaotic reality. Training large language models still commands headlines and keynote demos, but the real money has already migrated downstream to inference: the unglamorous, relentless work of serving predictions, generating tokens, and keeping chatbots responsive when millions of users hit them simultaneously.
Microsoft's Maia 200 accelerator, unveiled January 26, arrived purpose-built for inference with 216GB of HBM3e memory and the ability to cluster 6,144 chips in a single deployment. Google's TPU v5e emphasizes performance-per-dollar for inference workloads, a metrics shift that would have seemed mundane three years ago. Nvidia's Blackwell platform claims up to 25x inference cost reduction versus its predecessors, and the Rubin roadmap landing late this year leans even harder into long-context "reasoning AI."
The numbers tell a clearer story than the press releases. AI accelerator chips represented a $120.2 billion market in 2025. Forecasts place it at $433.3 billion by 2031, approaching $1 trillion by 2035—a 23.6% compound annual growth rate that has analysts using phrases like "once-in-a-generation." AI server shipments surged 46% year-over-year in 2024 and are expected to grow more than 20% again this year, reaching roughly 17% of total unit shipments but commanding 74% of server revenue. Alphabet is guiding toward $75–85 billion in 2025 capital expenditures, the vast majority servers and data centers. Meta's spending range sits at $66–72 billion, with signals the company intends to keep the spigot open through 2026.
That torrent of capital is chasing a specific problem: models are getting bigger, contexts are stretching longer, and agentic workflows—AI systems that reason through multi-step tasks—require memory-intensive operations that look nothing like the parallel matrix multiplication GPUs were designed to accelerate. Academic analyses and industry roadmaps both highlight what engineers call the "KV-cache swell" that occurs when models process extended contexts. Memory bandwidth, not raw compute throughput, becomes the constraint. GPUs optimized for training begin to stumble.
The Great HBM Squeeze
Nvidia's dominance—estimated between 70% and 90% of the AI accelerator market depending on whether you count only GPUs or fold in ASICs and FPGAs—rests on an integrated stack that rivals struggle to replicate: CUDA software that developers trust, NVLink interconnects that bind chips together, and privileged access to the industry's scarcest resources.
High-bandwidth memory remains the critical chokepoint, perhaps more than the company's founders initially expected. SK hynix signaled in March 2025 that its 2026 HBM supply is effectively sold out, with HBM3E 12-high stacks shipping primarily to Nvidia and HBM4 samples circulating to a select few customers. TSMC's CoWoS advanced packaging capacity—essential for bonding memory dies to compute chips—is expanding but remains tight. Monthly capacity is projected to reach 88,000–95,000 wafers by end of this year, up from roughly 75,000 in 2025, yet TSMC's CEO has noted publicly that advanced-node capacity remains "not enough" versus AI demand.
The supply-demand imbalance shapes everyone's strategy. Competitors who rely on the same HBM and CoWoS ecosystem face allocation risk and cost pressure that can derail roadmaps. AMD's MI300 series and the forthcoming MI400/MI450 Helios platform tout 50% more memory than Nvidia's Vera Rubin system, but they still lean on HBM and advanced packaging from the same bottlenecked suppliers. Intel's Gaudi 3 projects 50% faster inference throughput than Nvidia's H100 and emphasizes Ethernet-based scale-out to bypass proprietary interconnects—yet it too sources HBM externally. Microsoft's Maia 200 and Google's custom TPUs offer hyperscaler-controlled alternatives, but they compete for the same foundry and memory allocations.
This is the landscape Olix is explicitly designed to avoid.
The company's public "Compute Manifesto"—a document that reads part technical whitepaper, part philosophical treatise—describes what it calls a "contrarian belief" that scaling an SRAM architecture integrated with photonics can outperform HBM-based GPUs on throughput per megawatt and total cost of ownership. By rejecting HBM entirely and sidestepping CoWoS bottlenecks, Olix bets on different physics: light instead of electrons for moving data, on-chip SRAM instead of stacked memory towers.
Whether that wager pays off is the billion-dollar question—literally.
Light Moves Faster Than Electrons (In Theory)
Olix isn't alone in exploring optical approaches, though the sudden convergence of multiple well-funded photonics startups does suggest something real is happening beneath the surface. Lightmatter raised a $400 million Series D in October 2024, hitting a $4.4 billion valuation, with partnerships secured at Amkor and ASE for photonics-optimized 3D packaging. Celestial AI closed a $250 million Series C1 in March 2025, later confirming it's in acquisition talks with Marvell. Ayar Labs, working with Alchip on TSMC's COUPE optical I/O platform, has demonstrated 100 Tb/s target bandwidth per accelerator—a figure that makes electrical interconnects look quaint by comparison.
The appeal is straightforward enough, at least in theory. Photonics promises lower latency, higher bandwidth, and reduced power consumption for data movement between chips and across racks. As AI clusters scale from thousands to tens of thousands of accelerators, interconnect becomes the bottleneck—cables and switches can only move so many bits so quickly before physics intervenes. Nvidia's GB200 NVL72 rack packs 72 GPUs with NVLink fabric threading them together; AMD's Helios and Microsoft's Maia scale-up systems rely on high-speed Ethernet. Optical links offer a path beyond the power and distance limits of copper wiring.
But photonics is hard. Integration challenges, thermal management, manufacturing yield, and ecosystem maturity all remain open questions that startups must answer with working silicon, not PowerPoint slides. TSMC's COUPE platform and partnerships with packaging houses suggest the technology is moving from university research labs to pre-production, but volume deployment is still a 2026-and-beyond story. Olix's timeline—first customer shipments in 2027—acknowledges that reality without dwelling on it.
The company is hiring aggressively for optical compute roles and expanding its Bristol operations, tapping into the UK's South West and South Wales compound semiconductor clusters where firms like IQE and the CSA Catapult have built regional expertise. Whether that talent base can compete against Silicon Valley and Taiwan remains an open question.
The Inference Insurgency
Olix shares shelf space with a growing cohort of inference-focused challengers, each attacking different assumptions baked into GPU architectures. Groq's LPU—Language Processing Unit, a term the company trademarked—emphasizes deterministic, low-latency inference with public benchmarks showing 200–300+ tokens per second for Llama 70B models. Those are speeds that approach real-time conversation, the kind of responsiveness users notice. The architecture sacrifices flexibility for predictability, a trade-off that resonates in latency-sensitive applications but potentially limits the range of workloads it can handle efficiently.
d-Matrix, which raised $275 million in November 2025 at a $2 billion valuation, pursues what it calls digital in-memory compute (DIMC) with a 3D-DIMC roadmap that positions memory itself as the primary compute substrate. The pitch mirrors Olix's memory-centric thesis in spirit if not implementation: if inference workloads are memory-bound, redesign the chip so memory does the computation. Both startups argue that GPU memory hierarchies—optimized for training throughput over inference latency—can't simultaneously deliver high token rates and low cost for interactive AI.
Etched took perhaps the boldest stance with a $120 million Series A backing a transformer-only ASIC that hard-codes attention mechanisms directly into silicon. It's a bet that model architectures will stabilize enough that domain-specific acceleration can permanently outpace general-purpose chips. Hailo, meanwhile, targets the edge with its Hailo-10 GenAI accelerator, bringing large language models to PCs and automotive applications at under 5 watts—a $120 million April 2024 round underscoring investor appetite for inference efficiency at every scale.
What unites these efforts is conviction that Nvidia's training-optimized architecture leaves significant opportunity on the table for inference specialists. The metrics that matter have shifted: tokens per second and time-to-first-token for interactivity. Performance-per-dollar and performance-per-watt for sustained workloads. KV-cache capacity for long-context reasoning that extends beyond a few thousand tokens.
Groq positions on speed. d-Matrix on memory-centricity. Olix layers in photonics as a third axis of differentiation—assuming the physics cooperates.
The Execution Gauntlet
Raising $220 million buys runway, not success. That's an obvious point, perhaps, but one worth emphasizing given the wreckage of well-funded semiconductor startups scattered across recent history. Graphcore, the UK's last chip unicorn before Olix, burned through substantial venture funding and was acquired by SoftBank in July 2024 after struggling to compete against Nvidia's installed base and the gravitational pull of its ecosystem. Intel's own discrete GPU ambitions—years in development, billions in investment—have yet to capture meaningful share outside hyperscaler custom deals negotiated behind closed doors.
Olix must deliver working silicon, prove performance claims in customer production environments, and build software tooling that developers will actually adopt. All of this while Nvidia ships Blackwell systems today and Rubin platforms arrive in the second half of this year. The company's silence on customer partnerships as of February 2026 suggests either cautious stealth or early-stage sales conversations that haven't progressed to contracts. First silicon in 2027 means the market will have moved again—twice over—by the time Olix can prove itself in production.
The UK's semiconductor policy provides some tailwind, though perhaps less than boosters claim. The £1 billion UK Semiconductor Strategy announced in May 2023 prioritizes chip design, compound semiconductors, and R&D—areas where Olix and its photonics approach align with national priorities. Expanding in Bristol puts the company near former Graphcore talent and the South West's compound semiconductor ecosystem. Yet policy support doesn't substitute for market validation, and the UK lacks the domestic foundry capacity and packaging infrastructure that U.S. and Asian competitors access far more readily.
Regulatory dynamics could also shift in unexpected ways. U.S. export controls on AI chips and HBM, tightened repeatedly since October 2023, create complexity for any startup eyeing global markets. EU and French antitrust inquiries into Nvidia's CUDA bundling practices suggest possible openings for alternative platforms, but antitrust enforcement moves slowly and outcomes remain uncertain. Power constraints—Goldman Sachs forecasts a 165% increase in data center power demand by 2030, and the IEA projects 700–970 TWh of data center electricity consumption by 2035—may accelerate adoption of more efficient inference silicon. But only if performance and cost remain competitive, which is far from guaranteed.
The 2027 Moment of Truth
By the time Olix ships its first systems, the landscape will look substantially different than it does today. Nvidia's Rubin platform will be shipping in volume, with the company's now-recurring six-chip cadence (Vera CPU, Rubin GPU, NVLink 6, ConnectX-9, BlueField-4, Spectrum-6) setting a constantly moving target that competitors struggle to track. AMD's Helios racks will be deployed at Oracle and other early customers—or they won't, which would tell its own story. Microsoft's Maia 200 will have proven, or failed to prove, that hyperscaler custom silicon can displace third-party GPUs at scale in production workloads.
Analysts project Nvidia's market share declining modestly to roughly 67% by 2030 as Broadcom custom accelerators and AMD expand their footprint. That still leaves Nvidia commanding two-thirds of what could be a trillion-dollar market. For context: even losing a third of your market share when the market itself is growing 20%+ annually is a phenomenal business position.
The inference specialist cohort—Groq, d-Matrix, Etched, Olix, and others yet to emerge from stealth—will face their own collective reckoning. Which architectural bets paid off? Did photonics mature fast enough to matter? Did in-memory compute deliver on total-cost-of-ownership promises in real deployments? Did deterministic execution find a market niche large enough to sustain a standalone company? The venture capital deployed behind these companies—well over $1 billion across the cohort—reflects conviction that the answer is yes for at least some of them. But the semiconductor industry is littered with technically impressive chips that never found product-market fit, companies that solved the wrong problem or arrived at the right answer too late.
Olix's advantage, if you squint, might be youth itself. Dacombe is building with a clean sheet, unburdened by legacy architectures or obligations to installed bases. The company's manifesto emphasizes "model token hunger" and GPU memory limits for interactive inference—language that resonates with engineers debugging latency spikes in production chatbots and agent frameworks at 3 a.m. If that thesis is correct, and if photonics integration can be manufactured at scale without exotic yields, the timing could actually be right.
But the window is narrow and getting narrower. Inference economics are improving rapidly across all architectures as Nvidia, AMD, hyperscalers, and startups compete aggressively to reduce per-token costs. Power efficiency is climbing. Software frameworks are maturing. The question isn't whether inference accelerators will proliferate—TrendForce forecasts AI server shipments growing more than 20% annually through 2026, with ASIC share rising faster than GPUs—but whether purpose-built startups can carve out sustainable positions against incumbents with multi-billion-dollar R&D budgets and vertically integrated stacks that control everything from silicon to software.
Dacombe's $220 million buys the right to try.
The next eighteen months will determine whether Olix becomes the UK's next semiconductor success story—a narrative British policymakers desperately want—or another cautionary tale about the difficulty of displacing entrenched leaders in a capital-intensive, scale-driven industry where second place often means irrelevance. The company's bet on photonics and SRAM is bold, perhaps even inspired.
The market will decide if it's also right. It usually does.
