These sectors share a constraint: the science is ahead of the commercial path, and the winners are separated by execution rather than discovery. For each, what we think the real bottleneck is — and what we look for.
Embodied intelligence, perception, autonomy, and the manufacturing reality of putting models into moving hardware.
Not the model. The bottleneck is the gap between a demo that works in a controlled setting and a fleet that works in the world at a unit cost someone will pay. Reliability, service, spare parts, and the cost curve of the physical bill of materials decide this category — and almost none of that is a machine-learning problem.
Teams that have shipped hardware before, or have hired someone who has. A clear-eyed BOM with a credible path down the cost curve. A first application narrow enough to actually be reliable, rather than a platform pitch.
Semiconductors, edge and secure compute, novel architectures, and the frontier hardware the AI buildout actually runs on.
Design wins and the software around the silicon. A better architecture loses to a worse one with a mature toolchain, more often than anyone in this category wants to admit. The technical result is usually the easy part; the ecosystem is the moat and the years-long slog.
A realistic view of the design-win cycle and who has to say yes. Software and developer experience treated as a first-class product rather than an afterthought. Foundry, packaging, and supply relationships that already exist.
Computational biology, drug discovery platforms, brain–computer interfaces, and diagnostics where the bottleneck is engineering, not biology.
The business model, not the science. Platform companies in this category repeatedly discover that their customers want the output, not the platform — and that the regulatory and reimbursement path determines the timeline more than any technical milestone. Neurotech adds a hardware reliability problem on top.
A defined regulatory strategy with dates, not intentions. Clarity on whether the company is selling a tool, a service, or an asset — and the discipline to stop doing the other two. For neurotech, a serious answer on manufacturing and long-term device reliability.
Dual-use systems, autonomy, resilient communications, and the procurement path that decides whether good technology ever fields.
Procurement, and it is not close. The graveyard in this sector is full of technically excellent systems that never found a program of record, a budget line, or a sponsor willing to spend political capital. The timeline is measured in years and the buyer is an institution, not a person.
Someone on the team who has actually sold into this buyer and knows what a program of record requires. A commercial application that funds the company while the defense timeline plays out. Realistic ITAR, security, and clearance planning from the start rather than as a retrofit.
Power, cooling, interconnect, and the physical and energy constraints that now set the ceiling on compute.
Interconnection queues and power availability. Compute demand is no longer limited by chips; it is limited by megawatts, substations, water, and permits. That has turned an unglamorous infrastructure sector into one of the most consequential in technology — and the constraints are civil and regulatory, not computational.
Companies attacking a specific physical constraint rather than selling software about it. Credible utility and developer relationships. An honest read on the multi-year timelines that infrastructure actually runs on.
Security for AI systems, enterprise data protection, and the certification regimes that decide whether a deep-tech company can sell into regulated markets at all.
Two, depending on which side you are on. For security vendors: distribution and consolidation — platform vendors absorb good point solutions faster than in almost any other category, and a superior product with no distribution answer is a feature waiting to be acquired. For everyone else in these six sectors: compliance is a market-access problem wearing an IT costume. SOC 2, CMMC, HIPAA, and FedRAMP are not security exercises. They are the gate on the first enterprise, health-system, or government contract, and companies routinely discover the twelve-month timeline in month one of a sales cycle they expected to close in three.
For vendors, a clear answer to why this is a company and not a feature, and design-partner traction with buyers who hold budget. For everyone else, whether the certification path — SOC 2, CMMC Level 2, NIST 800-171, HIPAA, FedRAMP — is sequenced against the sales pipeline rather than started the week a customer asks for it.
Every one of these is a category where the principal has shipped product, led an acquisition, written a standard, or funded companies — usually more than one of those. Outside them we are reasonably informed readers, which is not worth paying for.
If your company sits between two of these, that is usually a good sign rather than a bad one. Most of the interesting work is at the seams.
Tell me what you are building, or what you are trying to underwrite. If it sits outside these six, we will say so quickly.