AI’s energy problem is a software problem
By Anna Trendewicz and Patrick Elftmann, Future Energy Ventures
AI’s energy problem is a software problem
The race to scale AI infrastructure is not an energy supply problem. It is a coordination problem, and coordination is a software problem.
Data center energy demand is set to more than double from 415 TWh in 2024 to roughly 945 TWh by 2030 [1], growing at more than four times the rate of total electricity demand growth. Interconnection queues in the US have swelled to 2,600 GW — twice the country’s total installed capacity, with median wait times approaching five years, and up to seven years in Northern Virginia. In the EU’s major markets, developers face waits of seven to ten years. A reported $156 billion in data center projects was blocked in 2025 alone [2]. Capital helps but it cannot fix permitting, sitting, or community opposition. The constraint has permanently shifted: the race is now about getting more out of what’s already connected. That is where software delivers and where we focus.
The constraint is not electrons — it’s intelligence
At Future Energy Ventures (FEV), we have spent over a decade building conviction that software is the most powerful lever in the energy system. Advanced economy power grids operate at just 30% average utilization2, not because power is scarce, but because the coordination systems to route it are outdated. According to Duke University, electricity providers can meet data center energy needs for 350 out of 365 days a year [3] — the constraint is 15 peak days, not baseload supply. GPU utilization in a typical enterprise AI cluster runs at only 30–60%. Google has reported doubling its data center energy efficiency driven primarily by software optimization. The 280x reduction in inference costs between 2022 and 2024 happened almost entirely through software improvements [4]. The conclusion is consistent: the biggest near-term efficiency gains do not require a new power plant, data center, or chip. They require better intelligence applied to the infrastructure that already exists.
Efficiency across layers also compounds rather than adds. A 10% gain at each of four layers delivers more value than a 40% gain at one. The most defensible companies will either span multiple layers or sit at the control point where tradeoffs are made.
Four layers of data center efficiency
Our deep dive structures the opportunity across four compounding layers. Together they represent a total addressable market of $251–362B by 2030.
1. Grid Efficiency — $12–24B by 2030
Software that navigates interconnection queues, turns data centers into flexible grid assets, and orchestrates distributed energy in real time. The core insight in this category is that advanced economy grids run at 30% average utilization because the coordination intelligence to use them efficiently doesn’t exist yet. This is a software and coordination problem masquerading as an energy supply problem, and it is the most underappreciated category in the stack.
Grid planning
Grid planning software helps data center and renewables developers navigate the interconnection queue through scenario modelling, capacity identification, and site selection. Physics-informed AI simulation tools are compressing weeks of engineering work into minutes, turning a multi-year bottleneck into a software problem. Key companies: ThinkLabs AI, PIQ Energy, Rhizome, Kevala, Neara, Paces, Pearl Street Technologies, LandGate, Blumen Systems.
Grid flexibility
Grid flexibility software aggregates demand response, virtual power plants, and battery trading to turn data centers into dispatchable loads that can absorb or shed power on command. This sub-category attracted the majority of grid efficiency VC investment through 2024, peaking at around $96M in a single year, before stalling as investors ran into the limits of near-term regulatory progress. Key companies: Emerald AI, Leap, Sympower, GridBeyond, Peak Power, Piclo, ev.energy, Gridmatic, Capalo AI, Enspired, Entrix, Suena, Flower.
Grid orchestration
Grid orchestration software treats the data center as an active grid participant, responding to price signals, dispatching flexible load in real time, and unlocking capacity without new infrastructure. The center of gravity in grid efficiency is shifting toward this sub-category. The business model innovation matters as much as the technology: companies that sell to data centers rather than utilities operate with 3–9 month sales cycles versus 12–24 months for utility-facing products and serve customers who are desperate for the product today. Key companies: GridCARE (FEV portfolio) recently closed a $64M Series A led by Sutter Hill Ventures. GridCARE’s physics-based AI unlocked 400 MW at Portland General Electric using pure software and zero new infrastructure, making it the largest orchestration round ever raised in this category. Other players include Camus Energy, Splight, Pebble, Mercury Computing, and Electron.
2. Facility Efficiency — $4–8B by 2030 (excl. hyperscalers)
Cooling optimization and infrastructure management software addresses the 40–60% of total data center energy costs attributable to thermal management. The key reframe here is that facility efficiency is capacity expansion, not cost reduction: every percentage point improvement in cooling creates additional sellable rack density without a new site. Investment was negligible from 2020 to 2024; 2025 marked the decisive inflection, driven by Phaidra’s $50M Series B and Gradyent’s $30M Series B.
Thermal management (cooling AI)
Reinforcement learning and AI are replacing rule-based cooling controls that have been largely unchanged for decades. The innovation frontier is multi-asset co-optimization: coordinating cooling, chillers, and compute workload simultaneously as a single system, rather than tuning each component independently. Digital twin technology enables full facility modelling for capacity planning and predictive failure avoidance. Key companies: Phaidra is the category leader, founded by Jim Gao, the engineer who led Google’s deployment of DeepMind’s AI-based cooling optimization. Its RL agent “Alfred” delivers 25% cooling energy reduction in production, with NVIDIA as a co-development partner. Etalytics is the leading European challenger, backed by Microsoft M12, with its etaONE platform delivering up to 50% HVAC and cooling energy reduction for customers including Equinix, Digital Realty, and NTT Data. Other players include Fluix, Ekkosense, Lucend, OctaiPipe, Monaire, Cosysense, and Zelena AI.
Waste heat recovery
Waste heat recovery software converts a cooling cost into a potential revenue stream by routing server waste heat into district heating networks, industrial processes, or on-site reuse. The technology requires real-time matching of heat supply from servers with demand from heating offtakers. The sub-category is nascent but structurally interesting as data center density continues to increase. Key companies: Gradyent, Qarnot, Energy Recombined.
Infrastructure management (DCIM)
DCIM software monitors power distribution, UPS systems, lighting, and physical infrastructure assets across a facility. The sub-category has produced no meaningful VC investment to date: incumbents Nlyte and WayLay were absorbed by Carrier [5] and Vertiv [6] respectively, and no new VC-backed challenger has emerged to replace them. The market remains dominated by bootstrapped players and legacy enterprise vendors. Key companies: Verdigris, Hyperview, Sunbird, Modius, Ethernetics.
3. Compute Efficiency — $60–100B in recoverable capex terms
Compute efficiency software recovers stranded GPU capacity and orchestrates workloads more intelligently across existing hardware. At 30–60% average GPU utilization across enterprise clusters, large share of installed capacity is sitting idle today, recoverable with software and no new infrastructure. NVIDIA’s behavior in this market is the clearest signal available: four acquisitions in 12 months (Run:ai [7], OctoAI [8], Deci AI [9], CentML [10]) confirm this as a strategic priority, as the company buys the software layer above its own hardware before someone else can.
Stranded capacity utilization
Stranded capacity utilization software identifies unused power and thermal headroom inside existing facilities and makes it deployable, enabling operators to safely run additional GPUs without new construction. The sub-category is almost completely unfunded, representing a sharp divergence from compute orchestration. Very few startups and very few investors are active here, yet the problem compounds in value every time a data center hits a power ceiling. Key companies: Hammerhead AI stands out with its ORCA platform, which uses multi-agent reinforcement learning across power, cooling, and compute simultaneously to deliver a 30% output boost within the existing power budget. The company is backed by SE Ventures (Schneider Electric) and was founded by AutoGrid veterans who bring grid intelligence expertise to the data center layer. Other players in this space include Utilidata, Neuralwatt, Pado AI, Bay Compute, Emerald AI, YellowDog, and Xperf.
Compute orchestration
Compute orchestration software intelligently routes AI workloads across GPU clusters, maximizing throughput, minimizing idle time, and scheduling jobs based on real-time energy cost and hardware availability. The sub-category is well-funded and increasingly mature. GPU utilization optimization for large clusters is now dominated by open-source tooling such as vLLM, Ray, and PyTorch, as well as hyperscaler-built solutions, which makes it a difficult environment for new startups to compete in. Investment accelerated dramatically in 2025, with Baseten and Modular now operating at late-stage scale. Key companies: Run:ai (acq. NVIDIA), OctoAI (acq. NVIDIA), CentML (acq. NVIDIA), MosaicML (acq. Databricks) [11], Modular, Anyscale, FlexAI, Expanso, Thunder Compute, Fireworks AI.
4. Software Efficiency — $175–230B by 2030
Software efficiency encompasses inference optimization, model compression, and novel model architectures. It is the most underfunded layer relative to its leverage. The 280x inference cost reduction between 2022 and 2024 happened almost entirely here. Investment was negligible from 2020 to 2023, at under $40M per year, then accelerated more than six-fold in 2025. Most category leaders have raised under $30M. The asymmetry between impact and funding is the opportunity.
Architecture efficiency
Architecture efficiency involves the fundamental redesign of model architecture to be more compute-efficient from the ground up, including state space models, mixture of experts, and neurosymbolic AI. The bet here is not on optimizing the current generation of models but on replacing the underlying engine entirely. Key companies: Cartesia is building state space models (SSMs) as a more efficient alternative to transformers, offering linear rather than quadratic scaling with context length. The company was founded by Albert Gu, the inventor of the Mamba architecture, and is backed by NVIDIA and Kleiner Perkins. Symbolica AI is developing neurosymbolic AI using category theory for structured reasoning, which is particularly relevant for industrial and energy applications that require auditability and rule-based logic. The company is backed by Khosla Ventures. Other players include Symbolic Mind, Elemental Cognition, and Mindbeam AI.
Model compression
Model compression reduces the size and compute footprint of an existing model before deployment through techniques including pruning, quantization, and distillation. The sub-category has already produced two significant exits, with Deci AI acquired by NVIDIA and Neural Magic acquired by Red Hat, leaving Pruna AI as the primary remaining independent pure-play. Strategic acquirer appetite is confirmed, but the pipeline of independent companies is thin. Key companies: Deci AI (acq. NVIDIA), Neural Magic (acq. Red Hat), Pruna AI, Doubleword.
Inference optimization
Inference optimization maximizes the throughput and minimizes the cost of running existing models on existing hardware, through techniques including continuous batching, KV cache management, quantization, and multi-model routing. Investment accelerated sharply in 2025 and 2026, driven by OpenRouter’s $113M Series B led by CapitalG in May 2026 and Inferact/vLLM’s $150M raise by a16z and Sequoia. The data moat from routing telemetry at scale, which reveals which models perform best for which tasks across trillions of tokens, is emerging as the primary competitive differentiator in this sub-category. Key companies: FriendliAI invented continuous batching through its Orca paper at OSDI 2022, which is now the foundational technique every inference engine runs on. Tensormesh commercializes LMCache, the leading open-source KV cache library with over 8,000 GitHub stars, already integrated into vLLM and NVIDIA Dynamo, and is backed by NVIDIA, AMD, and CoreWeave. OpenRouter operates a universal LLM API gateway routing across more than 400 model-provider combinations, processing 25 trillion tokens per week. Other players include Inferact/vLLM, Martian, NotDiamond, LiteLLM, and Refiant.
The control plane is the prize
The four layers are analytically useful, but the most defensible positions in this stack are not inside any single one. They are at the intersections. When a grid orchestration system can see real-time energy prices and that signal flows directly into the cooling system’s decision about how hard to run the chillers — which in turn determines how much thermal headroom exists for additional GPU load — the resulting optimization is worth multiples of what any single layer could deliver independently. A data center running at 40% GPU utilization, 30% above-baseline cooling cost, and retail grid pricing is not facing three separate problems. It is facing one problem with three levers. The company that controls the interface between those levers controls the economics of the whole facility. This is why the most interesting companies in this market are not the ones with the best single-layer algorithm. They are the ones building toward the control plane — where grid, facility, and compute decisions are made simultaneously — because that is where customers cannot afford to switch.
The exit market is already confirming the thesis
Exit activity to date has been concentrated in compute orchestration and model compression, the two most mature sub-categories in the stack. The acquisitions of Run:ai, OctoAI, and CentML by NVIDIA, MosaicML by Databricks, Deci AI by NVIDIA, and Neural Magic by Red Hat [12] represent a consistent pattern: the largest infrastructure companies in the world acquiring software efficiency capabilities they cannot build fast enough internally. NVIDIA alone made four acquisitions in adjacent categories within 12 months. The acquirer appetite is established, the strategic logic is clear, and the multiples reflect software valuations rather than infrastructure ones.
What the exit map also reveals is where the next wave will come from. Grid efficiency and software efficiency have produced almost no exits yet, not because the opportunity is smaller, but because these categories are earlier. Grid orchestration has only recently begun attracting meaningful capital, and software efficiency investment only inflected in 2025. Stranded capacity utilization has seen no meaningful exit at all. The acquirers are identified, active, and paying premium prices. The entry points in the less mature categories remain open.
FEV’s take
The efficiency layer is not a niche within energy transition investing. It is the precondition for the AI buildout. Data center energy demand doubles by 2030 regardless of which models win or which hyperscaler spends the most. The question is not whether that capacity gets built, it is whether the physical and financial infrastructure can support it. Software efficiency is the mechanism that makes it viable. That is the thesis, and we are building a portfolio around it.
Our edge is that we approach this market from the energy side rather than the compute side. Most investors arriving in data center efficiency are coming from infrastructure or AI — they understand the demand for compute but not the supply constraints at the grid and facility layer. FEV has spent years building conviction in grid intelligence, demand flexibility, and physical energy infrastructure as investment categories. We understand how grid operators think, how utilities procure software, how energy regulation creates both constraints and asymmetric opportunities for the right companies. That perspective is rare in a market where most capital is arriving from a compute-first worldview.
What we look for is precise. We want companies that unlock capacity from infrastructure that already exists, that sell to operators rather than waiting for regulators, and that sit at control points where decisions across grid, facility, and compute are made simultaneously. We are not looking for point solutions that optimize one metric in isolation. We are looking for the software layer that becomes embedded in how a data center operator runs its business and allocates capital — because those are the companies that are genuinely hard to displace.
Within this thesis, our current focus is grid orchestration, facility intelligence, and inference optimization: the three sub-categories where the ratio of impact to current funding is most asymmetric. We invest from late Seed through Series A, with initial checks of €1–10M, and a particular edge in helping companies navigate the European market, where the regulatory environment is complex, the data center buildout is accelerating.
If you are building in data center efficiency at the grid, facility, compute, or software layer, we want to hear from you. Contact Patrick Elftmann and Anna Trendewicz.
Sources
1 Energy and AI”, International Energy Agency, April 2025
2 AI doesn’t need more power, it needs a smarter grid”, World Economic Forum, March 2026
3 Rethinking Load Growth”, Nicholas Institute for Energy, Duke University, 2025
4 Stanford HAI 2025 AI Index Report
5 https://www.corporate.carrier.com/news/news-articles/202109_carrier-announces-agreement-acquire-nlyte-software-strengthen-expand-data-center-offerings.html
6 https://www.vertiv.com/en-us/about/news-and-insights/corporate-news/vertiv-acquires-generative-ai-software-leader-waylay-nv-to-enhance-critical-digital-infrastructure-operational-intelligence-optimization-and-services/
7 https://techcrunch.com/2024/12/30/nvidia-completes-acquisition-of-ai-infrastructure-startup-runai/
8 https://www.geekwire.com/2024/chip-giant-nvidia-acquires-octoai-a-seattle-startup-that-helps-companies-run-ai-models/
9 https://www.calcalistech.com/ctechnews/article/bkj6phggr
10 https://betakit.com/nvidia-acquires-canadian-ai-efficiency-startup-centml/
11 https://www.databricks.com/company/newsroom/press-releases/databricks-signs-definitive-agreement-acquire-mosaicml-leading-generative-ai-platform
12 https://www.redhat.com/en/about/press-releases/red-hat-acquire-neural-magic
