The Top AI Compute Startups in YC S26
The AI compute companies in Y Combinator's Summer 2026 batch are working on the cost of training and serving models. They sit at different layers of the stack, from power generation and silicon up through the models themselves, and several already have hardware running or customer workloads in production. Here are the top AI compute startups (in no particular order) from the batch.
Atomarine — Floating nuclear data centers
Atomarine (atomarine.co) builds standardized data center barges powered by nuclear ships, assembled on a single manufacturing line and towed to site. The pitch rests on what offshore siting removes: no interconnection queue, no competition with a town for water and electricity, and seawater cooling instead of an evaporative system. In its YC launch, the company points to interconnection timelines approaching a decade and to community opposition that has stopped a large share of planned onshore projects.
The founders are Dimitris Koutentakis, who holds four MIT degrees and has ties to the Greek shipping industry, and Emile Germonpre, who has a PhD in nuclear engineering. Atomarine says it has early partnerships with offshore barge designers and builders and with multiple small modular reactor vendors. The first pilot is planned as a gas-powered unit ready in 2028.
Orchestra — Distilling production traffic into owned models
Orchestra (orchestra.ai), which launched as Understudy, is an inference cloud that watches how a company's existing agents perform, turns those traces into evaluations, and trains smaller models to take over the work. A new model only gets production traffic once it beats the incumbent on the metrics the customer selected, whether that is task success, latency, cost or deployment size. Customers own the resulting weights and can run them in Orchestra's cloud, their own, or on-premise.
Luis Manrique and Aamir Poonawalla met building Instacart's optimization and experimentation systems. Manrique was later a founding member of technical staff at Gumloop; Poonawalla spent nine years at Instacart on ads serving and ML infrastructure and is a second-time YC founder. Their argument is that inference spend at most companies buys an answer and nothing else, while the traces that could train a cheaper specialist model get discarded.
Baud — AI chips without multipliers
Baud (baudlabs.ai) is designing silicon around a different representation of neural networks, one that removes multiplication from both the forward and backward passes and compresses weights by more than 10x. The trade-off is that models have to be trained in that representation, or start from a base model already converted to it. Because the cores never need multiplier circuits, they take less die area, which leaves more room for compute and SRAM.
Sarang Zambare, the CEO, was the ML lead for Peloton Guide and a founding engineer at Caper before it was acquired by Instacart. Eric Taylor, chief hardware architect, spent a decade in ASIC design with stops at NVIDIA, NXP, Arteris and Enfabrica. Baud's first chip is validated on GlobalFoundries' 12nm process and scheduled to tape out by the end of the year, and its training and inference service is already running on a cluster of FPGAs emulating the architecture. The team has built a compiler and distributed training stack that converts PyTorch-exportable models, including Qwen, Gemma and DeepSeek, into its format.
Experiential Labs — An open source gateway that trains your model
Experiential Labs (experientiallabs.ai) sits in front of whichever frontier models a company already uses and turns that traffic into a simulation of the tasks, failures and edge cases its agents run into. It trains a small language model against that simulation, routing more traffic to it as it improves while hard requests continue to the frontier model. Customers own the weights, and when a stronger open base model appears, the company says it carries the learned behavior over rather than restarting.
Kion Fallah, the CEO, earned an ML PhD at Georgia Tech and led the mixed reality simulation team at Waabi. Silen Naihin, the CTO, built and grew AutoGPT to 160,000 GitHub stars and previously co-founded Stackwise in YC's W24 batch. Experiential Labs says it has seen up to 97% lower cost on individual tasks, and it backs a 50% cost reduction at equal or better quality with an SLA.
Frontier Computing — Biological brains as a training substrate
Frontier Computing (frontier.site) grows neuronal tissue and sells access to it as a training and inference substrate. Academic work in the field has generally stalled around a million neurons because cells past a certain distance from an oxygen supply cannot survive, and Frontier says its culture approach gets around that limit. The appeal for ML is that memory and compute sit in the same place, which sidesteps the networking and bandwidth constraints that shape how large models are trained on GPUs. The company raised a $10M pre-seed led by General Catalyst, with LocalGlobe, Amino Collective, Kaya, Long Journey Ventures and others participating.
Michael Domarkas, the founder and CEO, is 19 and studied natural sciences at Cambridge; he has been culturing neuronal tissue for two years, starting with a $20,000 Emergent Ventures grant at 17. Frontier trained one of its systems to play Frogger, reporting a 92% peak crossing rate over a 25-game sample after an hour of real-time learning. It is building 100M and 500M neuron systems now, with the larger cluster targeted for January 2027, and plans an API for customers to run their own models and RL environments on the substrate.
OpenRelay — One endpoint across every accelerator
OpenRelay (openrelay.inc) takes an inference workload, schedules it onto whichever accelerator fits best, and hands back a single production endpoint. In its launch, the company describes routing across NVIDIA, TPU, Trainium and AMD hardware in more than four clouds, with load balancing, isolation, failover, metering and billing handled on its side, and claims cost savings of up to 20%. The network runs in both directions: operators with idle capacity, from data centers to reserved fleets to former mining clusters, can connect it and take inference demand without building a serving platform of their own.
Jaden Wang, the CEO, built a half-megawatt data center out of a warehouse shell at 20, then joined Voltage Park early to run virtualization and help stand up HPC; he previously worked at TensorDock and founded Heaviside Compute. Prashant Patel, the CTO, was a founding member of Amazon Bedrock at AWS, where he led Custom Model Import and wrote kernels to tune Trainium performance, and was later a staff engineer at Voltage Park working on inference and orchestration for clusters of more than 10,000 GPUs. OpenRelay says it is in production at 100 billion tokens a week, running on eight accelerator SKUs across 22 locations in Europe, Asia-Pacific, North America and the Middle East.