Beyond Software: Deep Tech Hardware Startups in Quantum, Photonic, and Neuromorphic Computing

Deep tech hardware attracted over $6 billion in venture investment globally in 2025, with quantum computing alone pulling in nearly $2.5 billion across more than 200 deals — figures that would have seemed fantastical a decade ago, when the prevailing wisdom said hardware was a graveyard for startups. That consensus is now dead. A new generation of companies building quantum processors, photonic chips, and neuromorphic computing architectures is drawing serious capital from sovereign wealth funds, defense agencies, and tier-one VCs who have concluded that the software layer is approaching its limits and that the next decade of performance gains must come from the physics of computation itself.

Why Hardware Is Back in Fashion

The renewed appetite for deep tech hardware has a straightforward cause: the AI boom has exposed the ceiling of conventional silicon. GPU clusters running transformer models at hyperscale consume megawatts of power and still cannot solve certain classes of problems — optimization at scale, drug molecule simulation, real-time inference at the network edge — without enormous cost. Software optimization can stretch existing architectures further, but the fundamental energy-per-operation cost of CMOS transistors has not improved meaningfully in years. Moore’s Law, as a practical economic force, is over.

Into that gap step three distinct paradigms. Quantum computing promises exponential speedups on specific problem classes using quantum mechanical superposition and entanglement. Photonic computing moves data on light instead of electrons, slashing energy consumption and latency in AI inference workloads. Neuromorphic chips mimic the sparse, event-driven firing of biological neurons, achieving orders-of-magnitude better energy efficiency for certain edge AI tasks. None of these is a drop-in replacement for GPUs or CPUs; each targets a specific class of problems where conventional approaches fail. But collectively, they represent the most significant architectural diversification in computing since the CPU itself.

Governments have accelerated the trend. The US CHIPS and Science Act, the EU Chips Act, and comparable programs in Japan, South Korea, and China have injected billions into semiconductor research and domestic fabrication. Defense agencies — DARPA in the US, DSTL in the UK — have funded quantum and neuromorphic programs for years. Academic spin-outs now move from lab to Series A faster than at any point in history, backed by a maturing ecosystem of deep tech accelerators and patient capital vehicles willing to hold positions across decade-long development cycles.

Quantum Computing: The Race for Practical Advantage

Quantum computing has the highest profile and the longest commercialization horizon of the three paradigms. The central question is no longer whether quantum computers can outperform classical systems on contrived benchmarks — they can — but when and on which real-world workloads they will deliver genuine commercial value.

IonQ, the publicly traded Maryland-based company, uses trapped-ion qubits — individual charged atoms held in electromagnetic fields and manipulated with lasers. Trapped-ion systems achieve significantly higher qubit fidelity than superconducting alternatives, making them attractive for near-term algorithms that cannot tolerate error. IonQ’s Forte system targets 35 algorithmic qubits and is available via AWS Braket, Azure Quantum, and Google Cloud, positioning the company as the cloud-native quantum option for enterprises beginning to experiment.

PsiQuantum, based in Palo Alto, is pursuing a radically different bet: a silicon photonics approach to fault-tolerant quantum computing that requires manufacturing at semiconductor fab scale. The company has raised over $700 million and partnered with GlobalFoundries to fabricate its photonic chips on commercial production lines. PsiQuantum’s thesis is that only a system with millions of physical qubits — far beyond any current device — will ever solve commercially relevant problems with sufficient reliability. The company is not building intermediate products; it is building toward a single, transformative machine.

QuEra Computing, a Harvard and MIT spin-out based in Boston, works with neutral atom arrays. Its Aquila system demonstrated 256-qubit operation in 2023 and has been made available via Amazon Web Services. Neutral atom platforms are notable for their ability to reconfigure qubit connectivity mid-computation — a flexibility that trapped-ion and superconducting systems lack. QuEra’s approach targets combinatorial optimization problems in logistics, materials science, and financial modeling.

The honest assessment: commercially relevant quantum advantage — meaning provably better results than the best classical algorithm at useful scale — has not yet been demonstrated for real-world business problems. Most quantum researchers place meaningful commercial deployment in the 2027–2032 window, contingent on progress in error correction.

Photonic Computing: Light-Speed AI Inference

While quantum computing promises future transformations, photonic computing is solving a problem that exists today: the energy and latency cost of matrix multiplication in large neural networks. Matrix multiplication — the core mathematical operation underlying every transformer, every convolutional neural network, every large language model — can be performed optically at the speed of light with near-zero energy dissipation, because photons do not generate heat the way electrons do.

Lightmatter, a Boston-based company founded by MIT researchers, has built a photonic AI accelerator called Passage that performs matrix math in the optical domain. Its second-generation architecture, Envise, is designed for data center deployment and positions directly against NVIDIA’s H-series GPUs for inference workloads. Lightmatter has raised over $400 million from investors including GV (Google Ventures) and Spark Capital, and has signed customer agreements with hyperscale cloud providers evaluating the technology for inference-at-scale deployments.

Ayar Labs attacks a related but distinct problem: the bandwidth bottleneck between chips. As AI accelerators grow more powerful, the copper interconnects linking them to memory and to each other become the limiting factor. Ayar Labs embeds optical I/O directly inside chip packages, replacing electrical signals with light for chip-to-chip communication. The result is dramatically higher bandwidth at lower power — an approach that works with existing processor architectures rather than replacing them. Ayar has raised over $130 million and counts Intel Capital, NVIDIA, and GlobalFoundries among its investors, a signal that the incumbent chip industry sees optical interconnects as inevitable.

The photonics sector has a meaningful near-term advantage over quantum: it does not require cryogenic cooling, exotic manufacturing processes, or entirely new software toolchains. Photonic AI accelerators can slot into existing data center infrastructure and run standard deep learning frameworks, lowering the barrier to enterprise adoption considerably.

Neuromorphic Chips: Computing Like a Brain

Neuromorphic computing takes inspiration from the architecture of biological neural networks rather than the physics of quantum mechanics or photons. Where conventional processors execute sequential instructions using billions of transistors switching billions of times per second, neuromorphic chips use spiking neural networks — sparse, event-driven circuits that only consume energy when a “neuron” fires. The result is extreme energy efficiency for tasks that map naturally onto sparse, temporal data: sensor processing, audio recognition, robotics, and always-on edge AI.

Intel’s Loihi 2, released in 2021, is the most advanced neuromorphic research chip currently available to external developers. Intel has distributed Loihi 2 systems to over 200 research institutions through its Intel Neuromorphic Research Community. Loihi 2 packs one million neurons and 120 million synapses into a 31-square-millimeter chip fabricated on Intel’s 4nm process, and consumes milliwatts — orders of magnitude less than a GPU performing equivalent inference. Intel has been clear that Loihi is a research platform, not a commercial product, and that a path to volume production depends on finding killer applications that justify custom silicon investment.

The Human Brain Project’s BrainScaleS system, developed at Heidelberg University, takes an analog approach. Instead of digital spiking circuits, BrainScaleS uses analog circuits that physically embody the differential equations governing neuron dynamics, running 1,000 times faster than biological real time. This makes the platform valuable for neuroscience simulation and for studying learning algorithms that are impractical to explore on digital hardware. BrainScaleS-2 wafers now support on-chip learning, a capability that most neuromorphic platforms still lack.

The commercial neuromorphic space is nascent. Startups including Innatera (Netherlands), SynSense (Switzerland), and BrainChip (Australia) are targeting ultra-low-power edge applications — hearing aids, industrial sensors, wildlife monitoring cameras — where battery life is the primary constraint. None has yet achieved the volumes that would make neuromorphic silicon a mainstream category, but edge AI applications are growing fast enough that the sector is attracting increasing attention from IoT hardware integrators.

Investor Landscape and the Funding Reality

Deep tech hardware investing requires a different model than software venture capital. Development timelines are measured in years and decades, not months. Capital requirements are enormous — a meaningful quantum computing program can consume hundreds of millions before a single commercial customer signs a contract. And the probability of complete technical failure is non-trivial; physics does not always cooperate with business plans.

The investor base has adapted accordingly. Sovereign wealth funds — particularly from the Gulf states, Singapore’s Temasek, and the Canadian pension funds — have become critical anchors in deep tech rounds because they have the capital scale and patience horizon that traditional VC structures cannot accommodate. Strategic investors from the semiconductor and defense industries provide not just capital but fabrication access, talent pipelines, and government contract relationships that pure financial investors cannot replicate.

Public markets have been a mixed experience. IonQ went public via SPAC in 2021 and has seen significant share price volatility as investors recalibrate timelines for quantum advantage. Rigetti Computing, another superconducting quantum company, similarly listed via SPAC and has struggled to maintain its valuation as commercial revenue growth has lagged technical milestones. The lesson is not that public quantum companies are uninvestable, but that the market is still learning how to price decade-long technology bets.

The most successful deep tech startups are those that have found intermediate commercial applications — products that generate revenue today from the same technology stack they are developing for the transformative application tomorrow. Ayar Labs selling optical I/O for conventional HPC clusters while building toward quantum interconnects is one example. Lightmatter running inference workloads on its photonic chips for cloud customers today, while the broader photonic computing vision matures, is another. The companies that struggle are those chasing a single moonshot with no near-term revenue bridge.

Sources & Further Reading

• IonQ — Quantum Computing Resources and Technical Documentation
• Lightmatter — Photonic AI Accelerator Technology Overview
• Intel Neuromorphic Research — Loihi 2 Program
• McKinsey — Quantum Technology Sees Record Investments, Progress on Talent Gap
• Ayar Labs — Optical I/O Technology
• QuEra Computing — Aquila Neutral Atom Quantum System