🌎 DeepSeek hype vs. the hyperscalers #232

Happy Monday! 

The Super Bowl might have got us hyped this weekend, and now, we’re taking on some other hype: the arrival of DeepSeek’s new AI model, and what its actual impacts will be on projected power demand. 

In deals, $45m for IoT-based building management, $36m for RNA-based technology for agriculture health  and $32m for alternative protein fermentation technology.

In other news, the Baltic States go energy independent, US-China tariffs take effect, and energy companies are cutting renewables investments. 

And don’t forget — take our readership survey to tell us what you want to see from CTVC going forward! We want to hear from you.

Thanks for reading!

Not a subscriber yet?

📩 Submit deals and announcements for the newsletter at [email protected].

💼 Find or share roles on our job board here.

DeepSeek’s (possible) deep cuts to power demand

For the past year, AI’s appetite for power has seemed insatiable — and hyperscalers were increasingly turning to clean firm power sources to meet it, until the arrival of the cheaper, more energy-efficient AI model DeepSeek called these assumptions into question late last month. Now, we’re taking a deep dive into this DeepSeek news to separate the hallucinations from reality.

What happened

On January 20, Chinese AI startup DeepSeek launched its R1 model, claiming it was far cheaper and more energy-efficient to train than US competitors. Costing just $6m to develop — compared to OpenAI’s reported $100m+ for GPT-4 — DeepSeek performed similarly to leading competitors while using only ~2,000 Nvidia H800 GPUs, an older chip and a fraction of what hyperscalers typically deploy. The model achieved this through two key differences in its training:

• Mixture of Experts (MoE) Architecture: Instead of updating all 671 billion parameters simultaneously like dense models, DeepSeek's system activates only a subset — just 37 billion per token — dramatically reducing computational overhead.
• Automated Reinforcement Learning: By minimizing human annotation needs, DeepSeek's model improves itself through immediate feedback, accelerating training while cutting costs.

The model also reportedly distilled competitors’ frontier models, raising the question of whether models can be easily replicated and distilled into smaller versions — which could potentially trigger LLM price wars and shrinking margins.

The market reaction was swift and severe across the AI infrastructure ecosystem. Poster child Nvidia lost 9% ($600B in market cap) before rebounding, and power utility stocks tied to data center growth — Vistra (-28%), NRG (-13%), and others — tumbled.

Why it matters

Data centers have driven rising electricity demand projections — pre-DeepSeek, the IEA projected that their consumption would more than double, hitting 650-1,050 TWh by 2026, largely from AI workloads. With today’s 8,000+ data centers already using 1-2% of global electricity, this trajectory required major infrastructure planning, which was already an open question. DeepSeek raises more new questions than answers.

There’s a difference between a general slowdown in power demand and a dip in hyperscaler demand. Utilities, with long planning timescales, were gearing up for large-scale data centers, each needing tens of megawatts of new capacity. Now, if AI models become more efficient, some workloads could shift to edge computing, reducing the need for massive centralized facilities, reducing the need for massive centralized facilities. DeepSeek’s breakthrough — cheaper, more efficient training — is just step one. Once trained, models move into inference mode for outputs. The better the training, the higher quality the outputs. But there’s a catch: efficiency gains in training don’t automatically mean lower energy demand overall. 

DeepSeek's energy savings may be smaller than assumed due to its higher per-query energy use and longer responses. Chain-of-thought outputs (like DeepSeek’s) consume 41% more energy than standard models — one complex response can require 17,800 joules, per MIT Tech Review, equivalent to streaming a 10-minute video. Since DeepSeek favors longer answers, its energy use can be up to 87% higher than concise models. Lower training demands may be offset by higher inference loads, even before Jevons Paradox — where increased efficiency drives greater use. In that case, an uptick in use would also create more distributed demand, reshaping how both tech companies and utilities plan the next generation of grid infrastructure. 

These developments underscore how difficult it is to accurately predict the amount of electricity demand needed in the next few years, and the concrete demand for clean firm power, especially novel solutions. Hyperscalers have been key buyers of nuclear, SMRs, and geothermal, supporting first-of-a-kind (FOAK) projects. If their energy needs decline, developers must find other major buyers.

What’s next

• Seeking tests for DeepSeek’s capabilities. Not all AI models are created equal. Early reports are that DeepSeek performs well in technical use cases, such as mathematical problems, reasoning or coding, but that US competitors are better with more creative requests. If DeepSeek is shown to have poorer performance in the fastest growing use cases, then its use could be limited and its impact on the energy system muted. 
• Stretching demand elasticity. Models that are cheaper to run should in theory be used more. However, researchers could simply run out of good data, hampering progress on training new models, or demand could naturally plateau, as it has for household electricity demand. Security concerns and potential bans on AI technologies could also mean uptake is limited.
• Hyperscalers’ next move: If major players like Microsoft, Google, Amazon, and Meta shift focus from aggressive power demand growth to competing with DeepSeek on lighter, more energy-efficient models, utilities may slow expansion plans — potentially slowing new generation buildout. SMRs face the biggest risk, as hyperscalers were expected to be key offtakers. Geothermal sits in the middle — data centers provide demand, but transmission investments allow broader supply. Large nuclear is least affected, able to serve both data centers and general grid expansions as AI usage grows.

Deals of the Week (2/3-2/9)

Late-Stage / Growth

🏠 75F, a Minneapolis, MN-based IoT building automation system provider, raised $45m in Series B funding from Accurant International, Breakthrough Energy Ventures, Carrier, Climate Investment, Copec WIND Ventures, and other investors. 

📦 Archive, a San Francisco, CA-based resale and circular fashion platform, raised $30m in Series B funding from Energize Capital, Bain Capital Ventures, Frontline, Lightspeed Venture Partners, and Woodline Partners.

Early-Stage

🥩 Liberation Labs, a Queens, NY-based precision alternative protein fermentation platform, raised $32m in Series A funding from Agronomics, Galloway Limited, Meach Cove Capital, NEOM Investment Fund, New Agrarian, and other investors. 

⚡ Presto Charging, an Oakland, CA-based EV charging platform, raised $15m in Seed funding from Congruent Ventures, Jetstream, Powerhouse Ventures, and Union Square Ventures. 

⚡ Tyba, an Oakland, CA-based solar and battery storage modeling platform, raised $14m in Series A funding from Energize Capital, Borusan Holding, Lorimer Ventures, Mobilize, Pear VC, and other investors. 

🌱 Edacious, a Marlborough, MA-based food verification system developer, raised $8m in Seed funding from Tin Shed Ventures, Jeremy and Hannelore Grantham Environmental Trust, Pelican Ag, The First Thirty, The Nest Family Office, and other investors.

🔋 ACT-ion, a Dallas, TX-based lithium ion battery cathode active materials developer, raised $8 million in Seed funding from BASF Venture Capital, Hunt Energy Enterprises, Mirae Asset Capital, Arosa Capital Management, and LG Technology Ventures.

♻️ MacroCycle Technologies, a Cambridge, MA plastics upcycler, raised $7m in seed funding from Clean Energy Ventures, Volta Circle and Indorama Ventures, KDT Ventures, and Neotribe Ventures.

⚡ Podero, a Vienna, Austria-based smart energy management platform, raised $6m in Seed funding from Planet A Ventures, PUSH Ventures, Pale Blue Dot, and Systemiq Capital. 

👕 Modern Synthesis, a London, England-based biotech textile developer, raised $6m in Seed funding from Extantia, Artesian VC, and Collaborative Fund. 

🌾 BiocSol, a Louvain-la-Neuve, Belgium-based microbial-free biopesticides developer, raised $5m in Seed funding from PYMWYMIC and Win4Company. 

🏠 Ampotech, a Singapore-based building energy IOT devices developer, raised an undisclosed amount in Seed funding from Clime Capital. 

Other

🧪 Greenlight Biosciences, a Boston, MA-based RNA-based sustainable agriculture technology developer, raised $36m in Debt funding from European Investment Bank (EIB). 

🥩 Liberation Labs, a Queens, NY-based precision alternative protein fermentation platform,  raised $19m in Convertible note funding. 

💨 MCi Carbon, a Mayfield West, Australia-based carbon utilization and storage technology provider, raised $5m in Corporate Strategic funding from Mitsubishi Corporation. 

🏭 Blastr Green Steel, a Lysaker, Norway-based low-carbon steel producer, raised an undisclosed amount in Corporate Strategic funding from Aurora Infrastructure, Cargill, Onvest, Security Trading, Tesi, and other investors.

New Funds

Hitachi, a Tokyo, Japan-based investment firm, launched $400m for their fourth corporate venture capital fund “HV Fund" to identify startups in digital tech (data centers, distributed energy, industrial AI) and emerging fields like biotech, quantum, nuclear fusion, and spaces. 

Overlap Holdings, a New York, NY-based investment firm, closed $33m for its first fund, Overlap Holdings Flagship Fund 1 (OHF1) which focuses on early-stage investments in energy, semiconductors, and related sectors.

Can’t get enough deals? See full listings and deal analytics on Sightline Climate.

In the News

The Baltic states—Lithuania, Latvia, and Estonia—disconnected from Russia’s power grid, marking a historic shift away from Soviet-era infrastructure as they prepare to integrate with the EU grid. The countries have prepared for this for the past two decades, and in the wake of Russia’s invasion of Ukraine in 2022 that sparked energy security concerns, brings a new era of energy independence, including potentially new renewables adoption, to the region.

In response to Trump’s recent tariffs on Chinese imports, China is imposing tariffs of 10% to 15% on US goods, including coal, liquefied natural gas, crude oil, and more, effective February 10. These retaliatory measures highlight the deepening trade conflict and its potential to disrupt global energy markets, with potentially broader effects on supply chains, pricing, and clean energy.

Equinor and Ørsted follow other energy companies in reducing investments in renewables. Equinor is halving its renewable energy investments from $10m to $5m and increasing oil and gas production by 10% in the next two years, while Ørsted, facing supply chain disruptions, permitting delays, and rising costs, is reducing its investment plans after setbacks in its US expansion. These moves reflect broader industry difficulties in scaling renewable energy projects amid a changing political landscape.

KORE Power canceled its $1bn Arizona battery plant, meant to be the first US-owned lithium-ion battery facility, citing corporate restructuring. Despite securing an $850m DOE loan, uncertainty around public funding and tariffs contributed to pauses in project development.

Gevo acquired Red Trail Energy’s ethanol production plant and carbon capture and sequestration assets for $210m, renaming it "Net-Zero North." With growing policy and industry support for sustainable aviation fuel, this move strengthens Gevo’s consolidation of the value chain for SAFs amid favorable market tailwinds.

In US agency news, several new department heads were confirmed, including former fracking executive Chris Wright as Secretary of Energy, oil and gas chief Doug Burgum as Interior Secretary, and rehire Neil Jacobs as NOAA Chief, who previously violated the agency’s code of ethics during SharpieGate. The EPA warned over 1,000 employees that they could be fired at any time to create an “effective and efficient” federal workforce.

In nuclear news, the UK is relaxing rules on new nuclear sites to fast-track construction for small modular nuclear reactors and make Britain a leader in SMR technology, while Belgium repealed a law mandating a nuclear phaseout and plans to revive its nuclear energy industry by extending the life of existing reactors and building new ones, aiming for a 4GW share in the country's electricity mix, amid a resurgence of momentum for the clean energy source.

Pop-up

It’s a bird! It’s a plane! It’s a low-emission flight technique inspired by migrating geese?

DAC is back: Climeworks dropped a guide to building and scaling Direct Air Capture plants.

A spotted hyena was spotted in Egypt, the first time in 5,000 years!

Former DOE LPO head Jigar Shah’s next move goes back to his roots with a new podcast launch.

Credit where it’s transferable: Crux put out an intelligence report for the transferable tax credit market.

The USAID shutdown removes global funding for renewable energy and disaster resilience.

Opportunities & Events

💡 Morgan Stanley Inclusive & Sustainable Ventures: Apply for the chance to join Morgan Stanley’s sustainability accelerators with a network of 80+ startups, offering $250K in funding, plus mentorship and investor access for underrepresented founders.

📅 Introducing Carbon Gap’s Policy Tracker 2.0: Register to attend Carbon Gap’s updated Policy Tracker launch on February 18th, featuring 500+ funding opportunities and programs, keynote speakers, and a panel of industry experts and policymakers. 

📅 College to Climate Career Center Program: Register to attend College to Climate’s Career Center Program on February 19th, designed to connect students with recruiters, provide information on opportunities, and give advice to young professionals entering the field.

💡 Climate Drift Career Accelerator: Apply for the chance to Climate Drift’s Career Accelerator, aimed at helping members build a network, solve real-world challenges, and transition into industry roles.

📅 UKRI Materials and Manufacturing Showcase: Register to attend the Materials and Manufacturing Showcase 2025 on February 19th— focusing on sustainable materials and manufacturing innovations, and featuring networking opportunities.

📅 Greentown Labs Climate Tech Intern Fair: College 2 Climate (C2C) is co-hosting an intern fair with Greentown Labs on February 27th and a prep session on February 24th to connect employers and job seekers across different academic backgrounds.

📅 Volta Foundation Battery Social: RSVP to attend Volta Foundation’s Battery Social on February 27th to connect with professionals and experts in the battery industry.

💡 Mark1: Apply for an opportunity to join Mark1, a Developer-as-a-Service for capital-intensive, FOAK industrial climate technologies, and maximize the potential for your technology to reach commercial scale.

Jobs

Research Associate – Grids, Senior Software Engineer, Marketing Manager, Customer Success Manager @Sightline Climate

Sales Lead, Head of Commercialization and GTM @Thalo Labs

Director of Business Development, Embedded / Full Stack Software Engineer, Senior Mechanical Engineer @Yard Stick PBC

Research Specialist, Economics & Finance @EFI Foundation

Founding Account Executive @CNaught

Full Stack AI Developer @Ezra Climate

Corporate Controller @Lunar Energy

Program Manager III, Commercial Partnerships, Advanced Clean Technologies @Google

Junior Scientist / Engineer, Senior Electrochemical Engineer @Oxylus Energy

Analyst @Alighned Climate Capital

Button: Find and share jobs here

📩 Feel free to send us deals, announcements, or anything else at [email protected]. Have a great week ahead!