New startup ideas · Physical world · Climate and energy
startup idea
Ballast Compute
GPU capacity API that runs on curtailed and behind-the-meter power worldwide.
Ballast deploys containerised GPU pods with onboard batteries at sites with power that cannot reach the grid: curtailed wind and solar farms, flared gas, oversized industrial connections.
- Infrastructure and APIs
- Enterprise
- $10-100B market
- Collapses a cost 10x
- Global from day one
4/5
venture judge
5
similar startups, last 2 years (19 all-time)
96%
of 4 nearest real companies still alive
yes
8 matching federal grants and programs
Direction supported by government programs and grants
Test it before you build it
$1,200 · 4 weeks · 25 prospects
For $1,200 and 4 weeks, prove that AI teams will prepay real money for interruptible GPU-hours at roughly $1 per H100-hour instead of demanding always-on capacity.
Riskiest assumption · ML teams at AI application companies have at least a quarter of their GPU spend in checkpointable fine-tuning and batch jobs, and will prepay for interruptible capacity at roughly $1 per H100-hour rather than insist on always-on GPUs, which is exactly what the stated risk and the industry operator's objection deny.
1Focus group: who and where
Head of ML infrastructure or CTO at a 20-100 person AI application company spending $25,000 or more a month on rented GPUs for fine-tuning and batch inference, whose compute bill grew faster than revenue this quarter and who is already shopping for anything cheaper than on-demand H100s.
where to find 25 · Start with the three AI application companies the founders already know and ask each for two introductions to peers with the same GPU bill. Post and DM in the MLOps Community Slack, where ML platform leads compare GPU vendors daily, and in the SkyPilot community Slack, whose users already run training on preemptible spot GPUs and have checkpointing working. Pull 100 AI companies from YC's public Startup Directory (S24 through W26 batches) and email the technical founder directly. Pitch in person at an AI Tinkerers meetup in San Francisco or New York, which run monthly and draw exactly these builders.
2Sell first, build later
A founding-cohort allocation of 5,000 interruptible H100 GPU-hours for fine-tuning and batch jobs, delivered through one API key with checkpoint and resume handled by the scheduler, live by February 1, 2027, capped at 20 customers.
the ask · $0.99 per H100 GPU-hour, against the $2.50 to $4.00 these teams pay on demand today; $2,000 refundable deposit per allocation, credited to the first invoice.
a real yes · A real yes is a paid $2,000 deposit or a signed pilot order with the February start date on it. Compliments, 'we would definitely use spare capacity', unpaid waitlist signups, and intros to someone else's infra team count for nothing.
3Small experiments
The first one attacks the riskiest assumption; each ends with a number that says whether to run the next.
1. GPU bill audit calls
$200 · 10 days
Book 15 twenty-minute calls with ML infra leads. On each call, walk through last month's actual GPU invoice line by line and classify spend into checkpointable work (fine-tuning, embeddings, batch inference, evals) versus latency-bound serving that cannot migrate, then pitch the founding allocation at $0.99 per H100-hour interruptible. A founder runs every call and logs the checkpointable percentage per company in a shared sheet.
keep going if · 8 of 15 audited buyers show at least 25% of monthly GPU spend in checkpointable jobs, and 5 of 15 ask to see the written pilot order
2. Priced reservation landing page
$700 · 10 days
Ship a one-page site with the $0.99 per H100-hour price, the interruption model stated plainly (jobs checkpoint, pause, and resume when site power returns), and a $2,000 refundable reservation checkout. Drive traffic with community posts, the audit-call follow-ups, and $500 of search ads on terms like 'cheap H100 hours' and 'spot GPU training'. Track visitors who reach pricing and who start checkout.
keep going if · Of the first 40 qualified visitors who reach the pricing section, at least 5 start the reservation checkout
3. Collect first-cohort deposits
$300 · 14 days
Send every warm buyer a one-page pilot order: 5,000 interruptible H100-hours at $0.99, first cohort live by February 1, 2027, $2,000 refundable deposit credited to the first invoice. Follow up twice over two weeks and record who actually pays versus who stalls on the interruption model or asks for an SLA the offer excludes.
keep going if · 5 of 20 pitched buyers pay the $2,000 deposit or sign the pilot order with a dated start
4Collect a deposit up front
Tesla took $1,000 refundable reservations for the Model 3 and $100 for the Cybertruck before building either: the deposit is the measurement, not the revenue.
$2,000
per prospect, refundable
how · A refundable reservation paid by card at the landing-page checkout, because this product's whole model is self-serve card-and-API-key signup, so paying by card now is the exact motion the buyer will repeat at launch; buyers spending over $50,000 a month sign the one-page pilot order and pay the same deposit by invoice instead, signed by whoever owns the GPU budget. set up: Stripe Invoicing ↗
what it reserves · One of 20 founding-cohort slots: 5,000 H100-hours at $0.99 locked for 12 months, priority scheduling at the first 2 MW site, and a named integration call to wire their checkpoint hooks before launch.
refund · Refunded in full, no questions asked, any time before first-cohort capacity goes live on February 1, 2027, or automatically if it does not.
target · 5 paid $2,000 deposits from 20 pitched buyers within 28 days
before taking money · This test sells compute reservations only; do not sign curtailed-power offtake, flared-gas, or behind-the-meter interconnection agreements before energy counsel reviews the state and FERC rules for each site.
Go: build it if
5 or more paid $2,000 deposits from 20 pitched buyers, and at least 8 of 15 audited GPU bills show 25% or more checkpointable spend: finance the first pod.
Kill: stop if
1 or fewer deposits after 20 pitches, or fewer than 4 of 15 audits show 25% checkpointable spend, or every interested buyer conditions payment on a non-interruptible SLA: the interruptible tier is the niche the risk section feared, stop.
5 Scripts to run itoutreach message, landing copy, deposit terms · click to open
outreach message
You're paying $2.50 to $4.00 per H100-hour for fine-tuning runs that already checkpoint every few minutes. I'm building Ballast Compute: interruptible GPU capacity sited on curtailed wind power, sold through one API at $0.99 per H100-hour, with checkpoint, migrate, and resume handled by our scheduler. Before we deploy the first 2 MW pod, I'm auditing GPU bills with 15 teams to measure what share of spend can tolerate preemption. Can I get 20 minutes this week to walk through yours?
landing page
Interruptible H100s at $0.99 an hour for checkpointable jobs A $2,000 refundable deposit reserves 5,000 H100-hours at $0.99, locked 12 months, live February 1, 2027 Reserve one of 20 founding slots
deposit terms
Your $2,000 deposit reserves one of 20 founding-cohort allocations: 5,000 interruptible H100 GPU-hours at $0.99 per hour, locked for 12 months, with capacity live by February 1, 2027. The deposit is credited in full to your first invoice. It is refundable any time before launch, and refunded automatically if we miss the February 1 date.
Would you run this test?
One tap. The yes-share feeds the Demand pillar of this idea's score; nobody sees who answered.
Budgets are out-of-pocket estimates for a team of one to three, US market. Size the deposit to the deal, and check the terms before taking money in a regulated line.
Scorecard
One score that balances how trendy the idea is, the demand for it and its potential for 100x, with competition measured relative to every other idea in the catalog. Recent startup trends first, government priorities second.
68
Idea Score, 0-100 · raw 42.2 x 1.61
Warm
competition: more crowded than 41% of ideas · headwind x0.80
+2.8
government priorities, secondary (17 matching grants)
Trend
46
Is the wave forming now? 2025-26 entrants vs 2023-24, rounds since 2025, the sector's live-batch direction, the 2026 trend analyst.
- Entrants 2025-26 vs 2023-24 (similar companies)13
- Rounds announced 2025+ in the sector47
- Sector direction (live batch)50
- 2026 trend analyst75
Demand
42
Does anyone want it? YC's current RFS, companies already paid for something similar, the operator judge, founders' yes-rate in decks, readers who would run the test.
- YC asks for it (current RFS: idea / sector)0
- Someone already pays (similar companies, recent / all-time)100
- Operator judge: real pain25
100x potential
63
Can it return a fund? The venture judge (double weight), market-size and moat axes, neighbours still alive, the technologist judge.
- Venture judge75
- Market size axis67
- Moat axis80
- Neighbours still alive6
- Technologist judge75
Score = 100 x cbrt(Trend x Demand x 100x) x (1 - 0.5 x crowding) + government bonus (max 5), calibrated so the 95th-percentile idea scores 90 (order never changes). A geometric mean: a weak pillar cannot be papered over. Percentiles are among the 272 ideas in the catalog; the terms matched were gpu, capacity, curtailed, behind-the-meter, power, worldwide, ballast, deploys.
The idea in full
- What
- Ballast deploys containerised GPU pods with onboard batteries at sites with power that cannot reach the grid: curtailed wind and solar farms, flared gas, oversized industrial connections. A single API sells interruptible training and batch inference capacity, and a scheduler built in-house checkpoints, migrates and re-prices jobs as power at each site appears and disappears. Enterprises sign up with a card and a key, and pay per GPU-hour at a rate set by the site, not by a datacentre rack.
- Why now
- The 2026 cohorts are full of the supply side of this trade and empty of the demand side: NODI Energy (YC X25) is building distributed energy infrastructure for data centers, Voxel Energy (YC W26) is doing energy independent data centers with solar and repurposed batteries, and voltiumAi and DataTherm are optimising and reusing data center energy, so the power is being assembled while no one sells it as an API.
- Wedge: first customer and entry point
- One 2 MW pod at a single curtailed wind site, sold as cheap interruptible fine-tuning capacity to three AI application companies the founders already know, on a raw scheduler with manual checkpoint hooks.
- Path to 100x
- Global AI compute spend is already a $10-100B annual market and stranded or curtailed power lets an interruptible tier price roughly an order of magnitude below on-demand rack capacity. The moat is distribution on both ends: multi-year exclusive rights to power at hundreds of otherwise unusable sites, plus being the default cheap tier inside the routers and orchestration frameworks that developers already call.
- Ceiling
- If grid power for data centers loosens or hyperscalers build their own behind-the-meter supply, the price gap compresses and Ballast becomes a low-margin reseller of hardware it financed.
- Closest real companies, as the generator saw them
- Voxel Energy and NODI Energy build the power and the buildings and sell space or electrons; Ballast never owns a building and sells compute-hours over an API, which is the layer above them. voltiumAi optimises energy inside existing data centers rather than placing workloads where power is stranded.
- Main risk
- Interruptible capacity turns out to be a niche because most enterprise AI spend is latency-bound inference that cannot tolerate migration.
Five judges
Each judge scores every idea in the catalog with a named rubric; the venture judge decides whether a card is shown at all (4-5 is venture-grade).
Venture investor
4/5
Interruptible GPU-hours at 10x below rack rates with exclusive multi-site power rights is a genuine $10B+ power-law bet.
Bootstrapper
1/5
A single 2 MW GPU pod with onboard batteries is tens of millions of financed hardware before the first billed GPU-hour.
Operator
2/5
Interruptible capacity forces buyers to rearchitect jobs around checkpointing and migration, which is exactly the workflow change enterprises refuse.
Technologist
4/5
An in-house checkpoint-and-migrate scheduler across intermittent 2 MW pods is genuinely hard systems work, though the moat leans on power contracts more than code.
Risk
2/5
Capital-heavy pods financed against interruptible demand from three known AI app companies, with site power rights as the single point of failure.
trends
4/5
Stranded-power GPU capacity matches the Compute at Sea RFS and AI infra's jump from 4% to 8% of YC S26; the window is open now.
Similar startups in the directory
Companies whose pitch matches most of the idea's terms (gpu, capacity, curtailed, behind-the-meter, power, worldwide, ballast, deploys): 19 all-time, 5 from the last two years. Same matching as Idea Check.
Microgrids in a box built anywhere in days
General Compute provides an ASIC-based cloud inference platform with an OpenAI-compatible API for deploying and serving AI models.
Kaspix helps OEMs ship cheaper, upgradeable edge devices faster—using swappable manufacturable analog-AI modules that cut BOM and test time without leading-edge digital chips.
Distributed energy infrastructure unlocking grid power to data centers
TROES is a Canadian company specializing in smart power energy storage technologies, products and solutions in a $170B market. We are an asset light tech company with systems powered by proprietary hardware & software tech which have been certified, commercialized and deployed globally.
Highly customizable and blazingly fast analytics for asset managers.
Data centers in space
ComfyUI for everyone on your team
The GPU Cloud Marketplace
Pitch Aeronautics
Acquired by Minu.mx - Fintech solutions for employees, starting with…
Voice AI Agents for customer engagement, including WhatsApp, SMS &…
The generator's reference companies
Real companies the model named as closest when it wrote the card, with their fate. A check mark is a company the radar could verify in its directory.
Public money in this direction
US federal grants, SBIR/STTR awards and open opportunities from the radar's public-money feed, matched to the idea's terms; the sector totals give the context.
8
grants and programs matching the idea
203
startup-relevant grants in Climate and energy
$167M
awarded in the sector, tracked
13
opportunities open now in the sector
- SBIR Phase II: In-Motion Capacitive Wireless Charging System for Material Handling Vehiclesawardhigh relevance
National Science Foundation · SBIR Phase II · $1M · posted 2026-09-03
- Development of a field-deployable tocoflexol-based countermeasure for multi-organ radiation mitigationawardhigh relevance
NIH / NIAID · SBIR phase I · $300K · posted 2026-09-01
- I-Corps: Translation Potential of an Ultra-Low-Power Microrobotic Sensing Platform for Deployable Distributed Robotic Systemsawardhigh relevance
National Science Foundation · I-Corps · $50K · posted 2026-08-17
- I-Corps: Translation potential of a compact power generation system based on rotating detonation combustion integrated with a turbine generatorawardhigh relevance
National Science Foundation · I-Corps · $50K · posted 2026-08-10
- SBIR Phase II: Stable Manipulation via Hovering Drone for Electric Utility Maintenance and Component Installationsawardhigh relevance
National Science Foundation · SBIR Phase II · $1M · posted 2026-08-04
- I-Corps: Translation Potential of a Multi-Modal Sensor Fusion Platform that Provides Precision Drone Swarm Trackingawardhigh relevance
National Science Foundation · I-Corps · $50K · posted 2026-08-04
- Developing Novel Ambroxol Analogs as Disease-Modifying Therapies for Alzheimer's Disease and Related Dementiasawardhigh relevance
NIH / NIA · SBIR phase I · $497K · posted 2025-09-16
- Pioneering continuous cell disruption for efficient commercial scale virus-like particle (VLP) biomanufacturingawardhigh relevance
NIH / NIAID · SBIR phase II · $1M · posted 2025-08-21
Market signal
What the radar sees in Climate and energy: new companies by cohort year, the forming YC batch, and outcomes since the February snapshot.
Climate and energy · 99 → 115 → 200 → 90 → 45 new companies 2022 → 2026 · 89% aliveYC F26 live: 1 in this cluster, 1% of the batch (was 1% in S26)Since February, of 118 YC companies here: 1 acquired, 2 shut down, 9 rewrote their pitch
Design attributes
The card is one cell of a designed set: every axis below was chosen before the text was written, and the text had to realize it.
- Buyer
- Enterprise
- Business model
- Infrastructure and APIs
- Path to 100x
- Collapses a cost 10x
- Market size
- $10-100B market
- Capital intensity
- Capital-heavy (hardware, bio, infra)
- Speed to revenue
- Revenue in 1-3 years
- Technical depth
- Deep tech: ML, hardware, bio
- Go-to-market
- Self-serve
- Moat
- Distribution moat
- Geography
- Global from day one
- Regulation
- Some regulation
- Vibe
- Hot space
Listed under
An idea sits in its own sector and in any sector its text clearly touches.
More ideas like this
Physical world · Climate and energy
Kilobid
Licensed AI agent that bids your home battery, EV and thermostat into wholesale power markets.
Kilobid is a consumer app that links a household's utility account, battery, EV charger and thermostat through existing APIs, then runs an agent that decides hour by hour when to charge, hold or discharge and submits those decisions into wholesale and utility program markets.
Physical world · Climate and energy
GridStake
The marketplace where grid interconnection rights are priced, traded, and transferred.
GridStake is a US marketplace where data center developers, renewable developers, and industrial buyers purchase interconnection queue positions and grid capacity rights from projects that will never build.
Physical world · Climate and energy
Ampsite
Marketplace where industrial sites sell spare grid capacity to data center developers.
Ampsite lists industrial and brownfield sites in the US that already hold energized interconnection capacity they no longer use, scores each one for real deliverable megawatts using utility queue data, feeder models and on-site meter history, and matches them to colocation and AI compute developers hunting for power.
Physical world · Climate and energy
Kilowise
Spend management for electricity - FinOps for every meter an enterprise owns.
Kilowise is self-serve SaaS that ingests utility bills, interval meter data, and tariff structures for every site a global enterprise operates, then finds billing errors, picks better tariffs, and forecasts power spend and carbon per facility.
Physical world · Climate and energy
Wattbid
The marketplace where small businesses sell their flexible power to the grid.
Wattbid enrolls the HVAC systems, EV chargers, and batteries of restaurants, gyms, and retail shops through the electrical and HVAC contractors who already service them, then auctions that aggregated flexibility to utilities and grid operators hungry for peak capacity.
Physical world · Climate and energy
Gridpad
Pre-permitted containerized power blocks that cut bridge power cost for data centers tenfold.
Gridpad manufactures standardized 2 to 5 MW containerized power blocks that combine battery strings, inverters and a grid-parallel controller, and ships each unit with the emissions permit and type certification already cleared for its site class in the jurisdictions Gridpad has pre-qualified.
Fictional company written 2026-08-22 from MarkosWeb data; the companies, grants and numbers around it are real and tracked. Treat the idea as a research prompt, not a plan.