GPU compute is expensive and orchestrating training jobs across cloud providers is a pain. TensorPool is a CLI that handles the infrastructure so ML engineers can stop being part-time DevOps.
GPU compute is the new commodity. Pluto is building a regulated derivatives exchange where you can hedge against compute price swings the same way airlines hedge jet fuel. The question is whether the market is ready to trade compute like oil.
Two former pilots are applying computer vision and AI to the problem of keeping planes from hitting each other. The FAA should be paying attention.
Every developer has wasted a week integrating an API that should have taken a day. ReJot wants to replace the SDK-plus-backend-wiring pattern with reusable building blocks that include the data layer. Two ex-Adyen engineers from Amsterdam who spent years thinking about distributed data.
FedRAMP compliance is the toll booth between software companies and the $100 billion federal IT market. It takes 16 months and costs over a million dollars. Archon says they can do it in six months and save you seven figures.
Potholes cost American drivers $3 billion a year. Pave Robotics is building autonomous robots that seal cracks before they become craters, and they work around the clock.
AI agents are stuck inside chat windows. A1Base wants to give them phone numbers, email addresses, and WhatsApp accounts so they can operate in the real world.
Most AI companies are obsessed with text and images. Overstand Labs is building foundational models for audio: speech-to-text, speaker diarization, and classification at scale.
Wiz sold to Google for $32 billion. SubImage thinks the same problem can be solved with open-source foundations and without the vendor lock-in. They might be right.
Most AI agent tasks are repetitive. Butter records tool call trajectories and replays them deterministically, cutting out the LLM entirely for cases that could have been a script. Faster, cheaper, and more predictable.
Most EV startups want to build the car. Steinmetz wants to build the thing that makes the motor spin, and they claim 3x the power density of existing solutions.
AI agents are shipping to production faster than teams can debug them. Subtrace traces every HTTP request and tool call without touching your code, giving you a real-time window into what your agents are doing and why they fail.
Every developer building AI agents is solving the same infrastructure problems from scratch. Truffle wants to hand them a solved stack so they can focus on the actual agent logic.
Kafka is powerful, popular, and notoriously difficult to work with. Quix built a platform that lets engineering and data teams develop, deploy, and monitor stream processing pipelines without wanting to throw their laptops out the window.
Every AI agent needs to talk to the outside world. Klavis AI is building the managed infrastructure layer so developers don't have to wire up OAuth flows and tool integrations from scratch.
SQLite for the web is a solved problem with about five solutions. Bunny Database is betting its CDN roots make it the one that actually sticks.
Deploying a compliant application used to mean hiring a DevOps team and a compliance consultant. Aptible packages secure, compliant hosting into a platform that startups can actually set up themselves.
Every AI analytics tool promises to answer your data questions. Cube says the others are lying to you — and it has 19,000 GitHub stars worth of infrastructure to make that case.
Knowing what software a company runs is only valuable if the data is fresh, sourced, and queryable — which is exactly what most technographics providers have historically failed to deliver.
Postgres scaling has always meant painful tradeoffs. PgDog, built by the engineer who sharded Postgres at Instacart during peak pandemic growth, wants to make horizontal scaling something you drop in rather than something you dread.
Power delivery hardware is boring until you realize data centers waste 30 TWh of electricity a year on it. PowerMatrix wants to shrink the problem by 80% with better semiconductor design.
Data centers run on five different software systems that don't talk to each other. Aravolta built one platform that replaces all of them and deploys in 48 hours.
Running GPU servers for AI models is expensive when they're busy and wasteful when they're idle. Cerebrium thinks serverless is the answer, and the economics actually make sense.
Background jobs are the plumbing of every serious application. Trigger.dev thinks the plumbing should be TypeScript-native, open source, and way less annoying to debug.
AI agents are shipping faster than anyone can test them. Janus is building simulation environments that catch the failures before they reach production. The pitch: trust, but verify.
Morphik is an open-source retrieval platform that gives AI apps access to real documents instead of confident fabrications. Space-tech teams and brokerage developers are already using it.
Two Berkeley dropouts built autonomous agent runtimes that get better at their jobs without human intervention. Their first agent, Rex, already handles GEO for 400+ customers including Rippling and Airwallex.
Data centers are expanding faster than operators can staff them. Boost Robotics thinks autonomous robots that can actually manipulate hardware are the only way to keep up with the inspection backlog.
Everyone knows RLHF makes models better. Almost nobody has the infrastructure to do it themselves. RunRL is betting that turning reinforcement learning into a managed service is the right abstraction.
Computer-use agents need somewhere to work that is not your laptop. Cua provides sandboxed cloud desktops where AI agents can see screens, click buttons, and run applications without touching your actual machine.
AI agents are calling APIs, querying databases, and executing transactions with long-lived API keys and zero oversight. Alter intercepts every tool call an AI agent makes and enforces parameter-level authorization, ephemeral credentials, and real-time guardrails. Two founders in New York are building the identity layer that the AI agent ecosystem forgot to build.
GitHub has become the default. That is the problem. gitdot is not building a GitHub clone with more features. They are building an opinionated platform for open-source maintainers who care about code quality over vanity metrics. No AI copilot. No free private repos. No stars. A Rust-powered Git server and a CI/CD platform designed for people who treat code as craft.
Browser automation breaks constantly because websites change. Notte built a full-stack browser infrastructure platform that lets AI agents run on the internet at production scale, with under 50ms latency, 1000+ concurrent sessions, and an agent success rate above 90 percent. Two MIT researchers who spent years studying why web automation fails think they finally fixed it.
Production debugging is still mostly manual. An engineer gets paged, opens Datadog, stares at logs, traces the issue across three services, and writes a fix four hours later. TraceRoot built an open-source AI agent that connects to your telemetry, traces the root cause, and drafts the pull request. The SDK has over 10,000 downloads and the founding team fixed 300+ production bugs at Meta and AWS before deciding to automate themselves out of a job.
The Model Context Protocol is becoming the standard way AI agents connect to the outside world. Manufact, formerly mcp-use, built the open source SDK that 4,000 companies already depend on. Now they are building the cloud infrastructure layer on top. With NASA, NVIDIA, and SAP as customers and 5 million downloads, this three-person team from Zurich and San Francisco is positioning itself as the default MCP platform.
Billing logic for AI products is a special kind of painful. Usage-based pricing, credit systems, feature gating, plan migrations. Autumn puts all of it in a layer over Stripe so you stop building billing code from scratch.
Every major AI model today requires a data center to run. DeepGrove is building technology to bring frontier-level intelligence to edge devices, from phones to embedded systems, without sacrificing capability. If they pull it off, the implications for privacy, latency, and accessibility are enormous.
AI coding agents are good at writing code but terrible at knowing what libraries actually do, what APIs have changed, and what documentation says right now. Nozomio indexes external context so your agents stop hallucinating about frameworks they last saw in training data.
Everyone wants to build AI agents that use computers like humans. The problem is you cannot train those agents on production systems without breaking things. Halluminate builds realistic sandbox environments that replicate Salesforce, Slack, and enterprise tools so AI labs can train and benchmark computer use agents safely.
Everyone is fighting over which model to run. Luminal is fighting over how fast you can run any of them. Their ahead-of-time compiler turns AI models into optimized GPU code and is already beating vLLM and TensorRT-LLM on throughput benchmarks. Three people, $5.3 million, and a very different theory of how inference should work.
AI agents are shipping into production with no way to safely test them against real-world failures. Playgent builds high-fidelity sandbox environments where teams can reproduce bugs, run evaluations, and train agents through reinforcement learning, all without touching live systems.
Coding agents are getting good at writing code. They are terrible at deploying it safely. Proliferate gives them sandboxed environments, event triggers, and verification workflows so they can ship without human babysitting.
Most companies running GPU clusters have 30 to 50 percent of their capacity sitting idle at any given time. Lilac unifies scattered GPU resources across clouds and on-premise into a single compute fabric, letting data scientists submit jobs to one scheduler instead of hunting for available machines.
Cloudflare Tunnels, Zscaler ZPA, and Ngrok all let you expose internal services to the internet. Pangolin does the same thing, built on WireGuard, fully self-hosted, with identity-based access controls and 19,000 GitHub stars to prove people want it.
AI agents fail in ways that are hard to reproduce and harder to diagnose. Fulcrum builds red-teaming agents that find the failures for you, uncover environment bugs, and catch reward hacking before it ships.
Everyone is building AI agents. Almost nobody can test them properly before deployment. AgentHub provides the simulation environments, tracing tools, and evaluation frameworks that agent developers need.
GPU data centers are burning through power at unprecedented rates and most operators have no idea how much they are wasting. DeepAware AI uses reinforcement learning to optimize workload placement and cut energy waste by up to 30%.
Traditional computer vision requires months of labeling data and training custom models. OnDeck uses vision language models to analyze video footage out of the box, no training needed, and their NeurIPS paper says it works better.
Every developer has built a weekend project that works locally but would take three weeks to make production-ready. Modelence ships everything you need to skip the boilerplate and go live from day one.
Mining companies and civil engineering firms sit on mountains of scanned maps, plats, and surveys that nobody can query because they exist as PDFs and JPEGs. Monarcha uses vision AI to turn those documents into real spatial data in under ninety seconds.
Voice AI is eating the world, but the training data supply chain is still held together with duct tape and crowdsourced recordings of questionable provenance. Panels wants to fix that with 10,000 vetted contributors across 100+ countries.
Monitoring tools tell you something broke. Interfere tells you what broke, why it broke, and how to fix it before anyone files a ticket. That is a meaningfully different product.
AI models that generate speech and video need training data that sounds and looks like real humans. Liva AI collects it with consent, diversity, and production quality baked in from the start.
Every sales engineer has experienced the empty-state demo disaster. Eden generates production-quality synthetic data on demand so your product looks like it has actual customers in it, without touching production databases or spending a week building seed scripts.
Every AI company has a GPU performance problem. Most solve it by hiring expensive kernel engineers. Wafer automates the entire workflow, from profiling to patching, and claims 1.5 to 5x speedups without anyone touching CUDA.
Building an AI agent is getting easier. Deploying one into production, with persistence, scaling, safety guardrails, and monitoring, is still a multi-week engineering project. Arcten wants to compress that into minutes.
Every AI company needs human annotators but finding, vetting, and managing them is a nightmare. Fixpoint automates the entire pipeline from sourcing to quality control.
GPU costs are the single biggest line item for most AI labs. SF Tensor claims it can cut those costs by up to 80% by automatically sourcing cheap GPUs across providers and optimizing kernels for any hardware. Three brothers, all CS graduates before turning 20, are betting that vendor lock-in is the real bottleneck in AI research.
Coding agents can write code. They cannot deploy it. Specific gives agents a cloud platform where they can spin up databases, configure secrets, and ship to production without a human touching the infrastructure.
Every AI agent forgets everything between sessions. Hyperspell builds the memory layer that connects agents to Slack, Gmail, Notion, and Drive so they actually know who you are and what you are working on.
AI agents are about to start spending money. Locus is building the payment rails that let them do it safely, with spending limits, escrow, and audit trails. This is infrastructure that does not exist yet.
Voice agents and chatbots need fast retrieval. Like, really fast. Moss built a Rust and WebAssembly search runtime that runs locally on any device and returns results in under 10 milliseconds. No database server, no network hop, no excuses. Grammarly and HubSpot are already customers.
Building a data center or factory means navigating hundreds of permits across dozens of agencies. Nerviom automates the entire regulatory maze with AI, cutting timelines that used to take years down to weeks. The market is massive, the pain is real, and the timing could not be better.
If you have ever tried to run three coding agents in parallel on the same machine, you have hit the port clash wall. Coasts fixes that with localhost isolation per git worktree, and it is free.
Kubernetes incident response is a nightmare of runbooks, alert fatigue, and 3 AM pages. Kestrel AI deploys agents that find root causes and generate production-ready fixes automatically.
Kafka is overkill for most streaming use cases and everyone knows it. s2.dev treats streams as a storage primitive with a REST API, sub-50ms latency, and no cluster to manage.
Cleft embeds feature flags directly in your codebase. No third-party SDK, no external dashboard, no vendor lock-in. Just flags that live where your code lives.
AI agents are everywhere and nobody can tell you why they broke. Fernstone gives developers tracing, evals, and cost tracking for autonomous AI systems so you can actually debug the thing before your customer does.
Your CI pipeline is rebuilding code that has not changed, and you are paying for every wasted minute. Trace Machina's NativeLink is an open-source remote execution and caching system that handles over a billion requests a month for companies like Samsung.
Most AI agent platforms give you a chat window. Helix gives each agent a full GPU-accelerated 4K desktop with a browser, terminal, and filesystem. It is a fundamentally different architecture and I think it is the right one.
Business intelligence has been promising self-service analytics for fifteen years and mostly delivering dashboards that nobody updates. Scoop takes a different approach: connect your data warehouse, ask a question in plain English, and get an AI that actually investigates the answer.
Every company deploying AI agents will eventually ask who has access to what. Agentic Fabriq is building the answer before the question becomes a crisis.
Deploying a coding agent to run overnight sounds great until you realize there is nowhere good to put it. Terminal Use wants to fix that.
Healthcare data is fragmented across dozens of systems that do not talk to each other. Eos AI harmonizes it all into something clinics can actually use.
Power utilities spend tens of billions inspecting their own equipment. Voltair is building the drone fleet to do it autonomously.
Rubric AI builds the human and computational layer that converts expert judgment into training signals for frontier models, focusing on regulated industries like healthcare.
Sequence Markets helps crypto traders save millions in hidden slippage by sourcing liquidity across CEX and DEX venues in a single workflow.
Compresr provides an API that compresses LLM context up to 200x without losing what matters, cutting token costs and improving accuracy for agents and RAG systems.
Constellation Space builds ConstellationOS, an AI layer that predicts satellite link failures 3 to 5 minutes ahead of time and autonomously reroutes traffic with zero data loss.
Chamber orchestrates GPU infrastructure with AI agents that detect unhealthy nodes, forecast demand, and reallocate resources so ML teams can run 50% more workloads on the same hardware.
Piris Labs offers a full-stack inference service using proprietary photonic hardware to eliminate the AI data movement bottleneck, delivering high-performance inference at a fraction of GPU cluster costs.
Maven enables AI voice agents to collect payments over the phone through a single API call, handling card processing and PCI compliance across all payment gateways.
Unifold provides a developer-first API and SDK that lets any app accept on-chain deposits across Ethereum, Solana, Bitcoin, and every EVM chain with less than 10 lines of code.
Captain provides a straightforward API for AI agents to search through unstructured files with 95% accuracy and citations, indexing data from S3, SharePoint, and Google Drive at petabyte scale.
The Token Company provides a drop-in API that compresses LLM inputs by 66%, cutting costs by 3x and improving response times by up to 37% while maintaining or improving output quality.
Shofo aggregates public and private video sources into a continuously updating index of billions of videos, then cleans, searches, and labels them to build custom datasets for AI research labs.
Travo provides the most comprehensive real estate data set with instant access to rental comps, ownership, zoning, GIS data, and financial analysis for any property or parcel in the US.
IncidentFox autonomously investigates production incidents by querying logs, metrics, and deployment history, delivering root cause analysis and one-click remediation scripts that wait for human approval.
Sonarly connects to your monitoring stack and turns noisy alerts into root-cause analysis and ready-to-merge pull requests. It is trying to be the on-call engineer that never sleeps.
Polymath builds complex multi-tool environments for training AI agents through reinforcement learning. Their Horizon-SWE benchmark tests what happens when you ask AI to do real work across real tools over long time horizons.
Voxel Energy builds energy-independent data centers that capture, store, and consume solar energy onsite using second-life EV batteries and a novel DC microgrid architecture, bypassing utility grid connection delays entirely.
Librar Labs is building self-healing data infrastructure that turns fragmented library records into agent-ready systems. Their school library product reportedly doubles student reading rates.
AI can see and understand the physical world, but building real-time vision applications is painful. Overshoot provides the infrastructure so developers can build security, safety, robotics, and consumer vision products without reinventing the stack.
Ashr generates synthetic user journeys across voice, text, images, and tool calls to test AI agents before they hit production. Prompt versioning with inline diffs shows you exactly what changed and why it broke.
Byteport builds file transfer infrastructure for companies that move massive files. Their DART protocol runs 10x faster than TCP for uploads ranging from 1GB to 100TB, serving robotics, satellite, AI, and enterprise applications.
AI agents need to spend money. Sponge lets them hold funds and transact autonomously using bank accounts, cards, and crypto, with budget limits and domain whitelisting to keep things under control.
Most AI training data has clear right answers. Traverse focuses on the hard part: capturing how humans make subjective decisions in law, healthcare, sales, and writing to train models that need judgment.
21st.dev started as the largest React component registry for AI apps with 1.4M developers. Now they are building a full SDK for deploying AI agents with built-in UI, memory, observability, and sandboxed execution.
RunAnywhere provides an SDK for deploying AI models on iOS, Android, and edge devices. Their MetalRT engine runs 1.67x faster than llama.cpp on Apple Silicon. 10K+ GitHub stars and growing.
APIs were not built for agents. Orthogonal gives developers and AI agents instant access to hundreds of APIs through MCP or SDK with pay-as-you-go billing. No API key management, no billing headaches.
Building a short-form video feed used to require a team of 20 engineers and a year of development. ShortKit gives you the same capability as a drop-in SDK with ML-driven pre-fetching, enterprise-grade transcoding, and native ad integrations.
Cumulus Labs is a serverless GPU inference provider claiming 12.5-second cold starts and scale-to-zero pricing, positioning itself as faster than Modal and cheaper than RunPod for teams running AI models in production.
Mendral is an AI DevOps engineer built by the people who wrote Docker's first lines of code, and it automatically diagnoses CI failures, catches flaky tests, and opens pull requests with fixes before your team even sees the red build.
Salus is an API that wraps around AI agents to inspect and block incorrect actions before they execute, preventing the kind of costly mistakes that evals and observability tools only catch after the damage is done.
Squid is building a browser-based workspace for electricity grid planning that replaces the spreadsheets and legacy tools utility companies use to model, stress-test, and plan the power networks that keep everything running.
Datost is an AI data analyst that lives in Slack, connects to your databases and data lakes, and lets anyone on the team query, debug, and analyze data right where they already work, without writing SQL or waiting for the data team.
Arga Labs provides sandbox infrastructure with fully functional mocks of external services so you can validate code changes and AI agent behavior at scale without risking production data or hitting real APIs.
Primitive is building the communication infrastructure that fully autonomous agents need to interact with the outside world, handling the email, messaging, and notification layers that agents cannot currently access on their own.
Minicor builds self-healing desktop automation that uses AI computer-use agents to operate legacy software like EHRs and ERPs that have no API, running on Windows VMs with video replay and automatic adaptation when UIs change.
The pitch is simple and kind of unsettling: let an AI observe everything your team does, find the patterns you never noticed, and start automating them.
