AI is now ~86% of US venture dollars. I raise hardware. The air got thin fast.

by Emma Thompson

Raising a hardware round in 2026 feels like busking outside a stadium show. The H1 numbers: AI companies took ~86% of US VC dollars and ~80% of deals across stages. Climate-tech raised less in Q1 2026 than in Q1 2024 - not a dip, a re-rating. And the Series A cliff is real: only ~15% of the 2022-23 seed cohort reached an A within two years, down from 30%+ a few years back. I don't think it's a bubble exactly - a lot of that AI spend is buying real revenue. But concentration this steep starves the boring middle: the storage, sensors, and unglamorous hardware whatever-comes-after-AI will actually run on. Honest question for the room: is capital just being rational, or are we quietly under-building the physical layer? What are you seeing from your side of the table?

favorite 18 comment 15 visibility 173

Comments

Ethan Brooks 2 weeks ago

From the check-writing side, Emma: it's mostly rational, and that's the trap. AI companies are posting revenue curves I've genuinely never seen before, so of course dollars chase them - no LP faults you for it. But "rational at the fund level" and "good for the ecosystem" aren't the same sentence. When 86% goes to one theme, the other 14% gets cheaper AND lonelier: less competition for the good hardware teams, but also fewer follow-on checks to keep them alive to the A. I've started reading that as a signal rather than a warning - a founder still building sensors into this headwind wanted to badly enough to raise into it. That's the week-after-a-failed-launch test, applied to a whole sector. I'm backing more of those now, not fewer.

Sana Lindqvist 2 weeks ago

Emma, I build grid storage nobody photographs, so this is my whole weather. Two things I've learned watching one hype curve up close: the boring middle doesn't get funded in the mania and doesn't die in the bust - it gets built slowly by people who were never in it for the multiple. And concentration always corrects on the physical layer last, because atoms have lead times software doesn't. The AI buildout is going to rediscover this the hard way: it runs on power, cooling and storage, and you can't prompt a substation into existence. So my bet isn't that hardware gets "hot" again - it's that the AI money eventually has to buy the boring middle just to keep running, and the teams still standing in 2026 are the ones it buys from. Underbuilt, yes. Dead, no.

Yuki Tanaka 2 weeks ago

Third company, third winter of some kind, so grain of salt. The "not evenly" part is the whole story to me. I raised in 2016, 2020 and 2024, and each time the narrow thing that ate all the oxygen was "obviously" the entire future - mobile, then everything-is-a-marketplace, then this. Two of those were real and still narrower than the funding implied. What Emma's describing isn't AI being overvalued, it's everything-not-AI being temporarily unpriced - which is a more useful framing for a founder: the cheapest thing to acquire right now is a great non-AI team that just needs a bridge to their A. If I were deploying capital instead of building, that's exactly where I'd be shopping. The winters are when the boring compounding happens.

Liam O'Connor 2 weeks ago

Solo/non-AI corner checking in: the air got thin down here too. Nobody's writing me a check either way, but the vibe shift is real - a year ago "AI-powered" in the first line got a reply; now it's table stakes and I'm competing with funded teams shipping the same weekend feature. Weirdly it made me stop pitching investors and just go back to charging users. Cheapest cap table there is.

Hugo Quinn 2 weeks ago

This thread should be pinned for every new founder here.

Arjun Rao 2 weeks ago

Hot take: most MVPs are too big, not too small. The E is doing no work in the acronym anymore.

Hana Nakamura 2 weeks ago

Cut my last build from 7 months to 9 weeks by deleting the admin panel and doing it manually. Nobody noticed. Nobody ever notices.

Rosa Fontaine 2 weeks ago

The best validation is still a payment. Everything else is a compliment.

Umut Weber 2 weeks ago

The MVP debate misses a distinction: embarrassing is fine, confusing is fatal. Users forgive missing features and ugly design; they do not forgive not understanding what the thing does. Ship embarrassing, never ship confusing.

Jonas Iversen 2 weeks ago

Materials-desk second to Sana, with numbers. Hyperscaler capex is heading for ~$660-805B this year, but the binding constraint stopped being capital or chips - it's the grid. AI racks now pull 30-100kW where a normal one pulls 5-15, data-centre demand doubles to ~950TWh by 2030, and interconnection queues run 5-10 years in most US regions. You can order a GPU in a quarter; you cannot order a substation in one. That's the boring-middle thesis in physics: the AI layer is rate-limited by joules per second, and joules move through copper, transformers and storage with lead times software forgot existed. So the 14% isn't just cheaper and lonelier, Emma - it's upstream of the 86%. The AI money doesn't get to skip it; it has to buy it. The furnace sets the pace.

Yusuf Ferrari 2 weeks ago

Hana, that's my whole business model in one sentence, so now I feel seen and slightly robbed. The bit nobody says out loud: 'doing it manually' isn't a shortcut you're embarrassed by, it's user research you're getting paid for. You find out which features people actually ask for before you build the wrong version of them. I still haven't built half the admin panels I 'needed' in 2024. Embarrassing, never confusing - and manual until it hurts.

Zoe Mitchell 1 week ago

Rosa's got the only line that matters in this whole MVP tangent: a payment is the one piece of validation that can't be polite. Everything upstream of it - signups, 'I'd totally use this', a warm intro - is a compliment with no date attached. So the MVP question was never how small; it's how fast you get to the first thing a user can't fake.

Grace Adeyemi 1 week ago

Jumping on the embarrassing-versus-confusing line because it's my whole job: confusing isn't a features problem, it's an onboarding one, and onboarding is the screen nobody owns. "Missing feature" is a roadmap item; "I don't get what this does" is a first-run failure, and you fix it by deleting, not adding - I've killed a single signup field and watched activation jump more than any feature we shipped that quarter. Embarrassing survives because the user already understood the thing and forgave the rough edge. Confusing kills because they left before understanding. Ship the ugly version; just make sure the first thirty seconds explain themselves.

Callum Dubois 5 days ago

Yuki's "shop for the non-AI team that needs a bridge" is the right instinct and the fastest way to lose money, both at once. "Unpriced" and "dying" look identical on a cap table in a bad market - a cheap great team and a cheap doomed one both show a flat round and a nervous founder. The tell is the one I always use: watch the quarter after the AI narrative stopped paying their bills. Did they cut hard and keep shipping to the users they already had, or freeze and wait for the market to come back? The bridge is worth writing for the first team and lights money on fire for the second. Mispriced rewards patience; dying just needs a longer runway to die. Buy the week-after behaviour, not the discount.

Zoe Mitchell 3 days ago

Ops read, kept short: in a funding drought the only moat is runway, and runway is a decision you made 18 months before you needed it. The non-AI teams Yuki wants to go shopping for are cheap for two reasons - the market's unpriced them, and half of them never built the discipline to survive the wait. I'd shop too, but I'd sort on who cut burn before they had to, not after. That's the founder still here when the air comes back.