Deep Tech Startups & Venture Capital: An Analysis of 2025 | Chapter 2
The annual report on the Deep Tech Cycle — 2025 in Data: 200+ hard numbers that matter, mapped month by month — Q3–Q4.
From fusion PPAs signed before electrons to humanoids-as-a-service and microreactors built on factory lines: the closing ledger of deep tech’s most industrial year so far.
By the second half of the year, the tape stops hinting and starts committing: power is pre-bought, robots are put on subscription, and sovereigns step in as LPs of last resort.
Greetings!
As you know, we believe that a sharper view of what’s ahead begins with a disciplined reading of what already happened.
In the first chapter of this annual, we traced 2025’s opening act — Q1 and Q2 through the numbers that mattered: advanced fission designs, AI megaprojects, sovereign funds, mine-waste platforms, and orbital logisticsmoving from slideware into concrete assets.
By the time we enter the second half of the year, the character of the tape changes. What showed up as prototypes, pilots, and program announcements in Q1–Q2 hardens into contracts, PPAs, factories, and sovereign capital structures in Q3–Q4. Fusion plants sign offtakes years before first electrons. Home batteries become dispatchable grid resources.
Microreactors move from concept art to manufacturing plans. Humanoid robots are priced as subscription opex, not one-off capex. GPU inventory is collateralized. EUV toolmakers, photonic interconnects, and diamond quantum foundries are treated as defense-grade infrastructure, not just semiconductor line items.
This chapter sits inside a four-part annual on the 2025 deep tech cycle.
The first half of the series—Chapter 1 and this Chapter 2—keeps the year in timeline mode, treating 2025 as a data object and following it month by month from Q1–Q2 to Q3–Q4.
The second half moves into theme mode, replaying 2025 through sector and thematic lenses—compute and AI infrastructure, energy and climate, critical materials and circularity, defense and space, and health and bio—to surface patterns, recombinations, and deeper dynamics.
Over these six months, several patterns crystallize:
Frontier AI and “physical AI” are capitalized as infrastructure.
Sovereign-leaning labs raise multi-billion-dollar tickets for open or open-weight models. Robotics-first foundation models and “robot brains” are financed as horizontal control layers that sit on third-party hardware. Humanoids are priced below $500 per month, and a projected $5T humanoid market by 2050 gives embodied autonomy a macro demand envelope on par with autos or smartphones.Energy and heat are wired directly into the AI stack.
Microreactors with five-year refueling cycles, heat-only SMRs, superhot-rock geothermal, “AI-grade” closed-loop projects, and fusion PPAs signed before electrons reposition firm power as a front-row constraint rather than an externality. Home batteries aggregate into virtual power plants, thermal batteries become heat-as-a-service, and state vehicles like Deutschlandfonds deliberately blur the line between energy policy and industrial strategy.Critical materials and circularity become programmable.
Metal-fuel generators, lithium refineries, recycled NdFeB magnet plants, microbial biomining, electrochemical DLE, automated de-manufacturing, and AI-designed substitutes for critical minerals all treat metals and feedstocks as a process-engineering and compute problem—not merely a question of finding new mines.Orbital and subsea domains turn into logistics and sensing layers.
Microgravity CDMO platforms, reusable microgravity labs, electromagnetic launchers, debris-tracking constellations, warehouse-style robots for space stations, air-breathing propulsion in VLEO, and swarms of autonomous underwater vehicles sketch a midstream of space and subsea infrastructure servicing pharma, defense, and climate models from orbits and oceans.Policy, compliance, and underwriting become the real bottlenecks.
EDF work programs, agricultural-robotics competitions, regenerative agriculture credit, grid-compliance automation, and lengthening venture fundraising cycles all point to the same reality: the speed limit is increasingly regulation, risk appetite, and institutional plumbing—not whether the core technology works.
What follows is the ledger of deep tech’s second half of 2025—Q3 to Q4—written in numbers.
Enjoy the read!
– The Scenarionist Team
July 2025 – Frontier AI Labs, Physical AI and Nuclear Heat Move into Strategic Capital Structures
Frontier AI labs and “physical AI” move into sovereign and infra-grade territory, with multi-billion-dollar funds and national-security vehicles treating open-source multimodal models, robotics-first AI, and photonic processors as strategic assets rather than speculative software bets. Superhot-rock geothermal and nuclear heat—via district-heating SMRs and microreactors with five-year refueling cycles—enter urban and remote-infrastructure planning, while methane pyrolysis, molten-oxide electrolysis, and lithium refineries reframe metals and hydrogen as process-industry decarbonization levers. In parallel, long-duration CDR offtakes, cultivated-meat megaplants, microgravity CDMO platforms, and sovereign-backed EO constellations make July a month where capital structures and offtake logic, more than headline valuations, show which technologies are being pulled onto state and corporate balance sheets.
$2.5B – Greenoaks’ $2.5B growth fund enlarges the global late-stage capital stack available for durable software, AI and infrastructure bets, reinforcing the role of concentrated crossover managers as key allocators into “winner-take-most” platform companies.
$2B – Thinking Machines Lab’s $2B seed round at a $12B valuation to build an open-source, multimodal frontier lab led by Mira Murati positions the company as a de facto “sovereign AI” asset, with capitalization closer to strategic infrastructure than to a conventional early-stage software venture.
$105M – Genesis AI’s $105M seed to build universal robotics foundation models based on large-scale simulation and synthetic data formalizes “robotics-first” AI as its own foundation-model category, with ambitions to become the neutral control plane for heterogeneous physical robots.
€62M – Q.ANT’s €62M Series A to commercialize thin-film lithium niobate photonic processors, promising up to 30× energy efficiency and 50× speed-ups for AI and HPC workloads, is one of the few AI hardware bets that targets the cost-per-token denominator rather than just adding more GPUs.
1,000 °C – the 1,000 °C rock temperatures reached in Quaise’s millimeter-wave drilling tests illustrate why superhot rock geothermal is treated as potential baseload: operating at rock-melting regimes that traditional drilling cannot tolerate, with a patent stack explicitly designed for >1,000 °C borehole environments.
$80M – Bedrock Robotics’ $80M Series A to retrofit excavators and bulldozers into autonomous platforms shifts “physical AI” from selling new robots to upgrading existing heavy equipment fleets, compressing payback periods and aligning with brownfield CAPEX cycles in construction and mining.
4.9M tCO₂ – Microsoft’s offtake for up to 4.9 million tonnes of CO2 removal from Vaulted Deep over 12 years, against roughly 15 Mt of annual emissions with >97% in Scope 3, reframes waste-management infrastructure as a long-duration CDR asset class rather than a pure compliance cost.
$51M – Boston Metal’s $51M convertible note to scale molten-oxide electrolysis for critical metals in Brazil while preparing a green steel demonstrator in the U.S. underscores MOE as a serious process route for decarbonizing iron and specialty metals, not just a research curiosity.
200,000 ft² – Believer Meats’ 200,000-square-foot cultivated chicken facility in Wilson, North Carolina, one of the first large-scale GMP plants cleared for U.S. sales, provides a concrete scale and capex reference for cell-ag plants even as the company’s business volatility highlights the sector’s execution risk.
$46M – Pure Lithium’s $46M investment to relocate HQ and pilot manufacturing to Chicago and vertically integrate lithium-metal anodes marks a transition from R&D into domestic industrialization, supporting North American battery resilience beyond conventional graphite-heavy supply chains.
$27M – Tulum Energy’s $27M raise for turquoise hydrogen using electric arc furnaces and solid carbon co-product advances methane pyrolysis from lab-scale into process-engineering territory, aiming for H₂ costs closer to SMR while generating a solid carbon stream that can plug into existing materials markets.
20,000 t/year – Mangrove Lithium’s planned refinery with 20,000 tonnes