Where we'd put infra money.

A data center can take delivery of GPUs in a quarter and then wait three to seven years for the grid connection that powers them. GridCARE argues the capacity often already exists, idle at the wrong hours and the wrong substations, and that the real work is finding it. Its Energize platform models the physical grid to locate that headroom and sells speed to power.

The proof is specific. A partnership with a utility in Hillsboro, Oregon clears a path the company puts at more than 400 MW, with the first 80 MW arriving in 2026. Sutter Hill led the round, an original Nvidia backer, and National Grid's venture arm joined. That cap table suggests utilities themselves take the model seriously.

If the grid is the gate, the wildest way through is to leave the grid behind. Orbital wants to run AI inference in low Earth orbit, where the sun never sets and the cold of space does the cooling. No interconnection queue, no water, no land fight. The founder is Euwyn Poon, who co-founded the scooter company Spin and sold it to Ford in 2018.

Temper it. This is the longest shot on the page by a wide margin. Launch cost, radiation, and on-orbit servicing are all unsolved at scale, and a $5 million pre-seed from a16z's Speedrun buys a single demonstrator rather than a data center. The first mission, Pathfinder, is slated to fly a hosted GPU payload on a Falcon 9 rideshare in 2027.

Inference spends much of its time and energy shuttling weights between memory and the processor. Fractile builds chips that do the math inside the memory itself, which is where its claim of running frontier models up to 100 times faster and 10 times cheaper comes from. Walter Goodwin founded it out of Oxford in 2022; this round pushed it into the unicorn bracket.

Forget the raise for a second; the buyer is the tell. Anthropic is reported to be in early talks to purchase Fractile's chips once they ship. A frontier lab shopping for inference silicon outside Nvidia is the demand this whole category is counting on, and a named customer before first silicon almost never happens.

Photonic computing has been five years away for a decade, because the optical parts were too big to pack tightly. Neurophos, spun out of Duke metamaterials research, says it shrank the optical modulator by roughly 10,000 times. That is enough to put on the order of a million optical elements on a single die and make the approach buildable. When it works, inference moves through the chip at light speed for a fraction of the energy.

The hard, ownable piece is the modulator itself; the packaging around it is the easy part. Gates Frontier led the $110 million round and Microsoft's M12 joined, the kind of patient money that funds a chip on a 2028 horizon. The team carries veterans from Nvidia, Intel, AMD, and Lightmatter.

Read this one as a founder purchase first. Naveen Rao built Nervana, which Intel bought for $350 million, then MosaicML, which Databricks bought for $1.3 billion. A $475 million seed at a $4.5 billion valuation for a two-month-old company, with Rao adding $10 million of his own, is the market pricing that record against the energy wall.

The technology is unproven. That is the honest description of pre-product analog silicon. The case is that the person who industrialized two earlier compute shifts has earned a third look, and the round gives him a decade of tape-outs to find it before revenue. a16z and Lightspeed co-led; Lux, DCVC, and Jeff Bezos are in.

Inside the rack, copper is running out of room: every doubling of bandwidth burns more power than the last. LightSpeed Photonics, founded in Bangalore in 2021, builds optical interconnects it says move data four times faster, at half the power, in a twentieth of the space. As clusters grow to tens of thousands of GPUs, the links between chips choke throughput as badly as the chips do.

pi Ventures led the $6.5 million pre-Series A, with about $8.5 million raised so far including grants. This is component-level deep tech that every hyperscaler needs and few early teams attempt. The pilots that count are the ones with OEM and ODM partners; those turn a lab number into a product.

The cleanest way to lean against Nvidia's pricing power is to fund the one other accelerator a serious lab will actually run. TensorWave operates one of the largest liquid-cooled AMD Instinct fleets, more than 8,000 MI325X parts, and reports a $100 million-plus run-rate growing 20x year over year. AMD's own venture arm co-led the round, which means the chipmaker is underwriting its cloud channel.

This is one of the rare infrastructure startups already at a nine-figure run-rate instead of a pitch deck. If AMD keeps narrowing the software gap to CUDA, TensorWave is positioned to sell the supply that gap frees up.

As agents multiply, buyers stop caring which GPU runs a model and start caring about price per token and time to first call. Parasail abstracts a fabric of compute and sends each job to the cheapest endpoint that still clears the latency bar, billed per token, live in minutes. Moving more than 500 billion tokens a day about a year after launch says the demand for that abstraction is real.

Sitting above the metal keeps Parasail off the hook for any single chip, and the agent wave is exactly the workload that makes token counts explode. The open question is margin: can the routing layer hold its cut while the labs squeeze from above and the neoclouds from below.

Most GPU clusters run at 30 to 50 percent utilization, which is paid-for hardware sitting dark. Lilac, a Y Combinator company from the summer 2025 batch, routes inference onto that idle capacity and pays the owner for it. The interface is OpenAI-compatible, the billing is per token, and the plumbing is Kubernetes-native. It is the supply-side mirror of Parasail's demand-side routing.

The founders, Ryan and Lucas Ewing, both come out of cloud infrastructure. Their pitch turns a stranded asset into supply, and OpenAI compatibility makes adoption a config change rather than a rewrite. Whether it works comes down to liquidity and reliability.

An agent is only as good as its reach into the live web, and the web was built for human eyes rather than API calls. Parallel sells the search and research APIs that let an agent read, reason over, and act on web data. Parag Agrawal, the former Twitter CEO, founded it; Clay, Harvey, Notion, and Opendoor are customers, and the company reports more than 100,000 developers.

A valuation that tripled to $2 billion in five months, with Sequoia leading the Series B, says investors think the agent layer needs its own infrastructure company. The sharper move is the plan to pay publishers when agents use their work. That is the unsolved economics of the agent era, and if Parallel makes it a standard it owns a toll booth.

AI broke the assumptions object storage was built on: billions of tiny files like embeddings and shards, read at latencies S3 was never tuned for, plus egress fees that punish moving data to wherever the GPUs are. Tigris is a globally distributed store that speaks the S3 API and is built for that access pattern. It has more than 4,000 customers and is expanding from the US into London, Frankfurt, and Singapore.

Compatibility is the wedge. Code written for Amazon S3 runs on Tigris unchanged, so the switching cost rounds to zero, and the offer underneath is lower latency plus relief from the cloud's most profitable charge. Spark Capital led the $25 million round, with a16z returning.

GPUs sit idle while data copies down from object storage. Archil, formerly Regatta Storage, mounts an S3 bucket as a POSIX filesystem that behaves like a local disk, so training and inference start before the download finishes. Full POSIX support means no code changes, the same near-zero switching cost that makes storage wedges work. The founder ran Amazon's Elastic File System and Netflix's cloud storage, the exact résumé this problem wants.

Felicis led the $6.7 million seed, with Y Combinator in and angels including Modal's Erik Bernhardsson and a former AWS storage lead. The risk lives next door to the wedge: this is close to something AWS could ship as a feature.

Custom silicon is the obvious hedge against renting Nvidia, and it still takes a large team and several years to tape out an ASIC. Architect Labs, out of stealth in June 2026, is building an AI system that designs and verifies a chip end to end. The founders are Ebrahim Hussain, who worked on silicon at Apple and Tesla, and Aaditya Subedi, a former Harvard researcher in code verification.

The cap table is the loudest signal here. Kindred Ventures led the $24 million seed, and the angels include Google's Jeff Dean, OpenAI's Lukasz Kaiser, Perplexity's Aravind Srinivas, and the RISC pioneer Kunle Olukotun. Those are the people who would know whether AI-designed silicon is real.

The next gains for many agents come less from a bigger base model than from teaching a smaller one your specific task. Osmosis, a Y Combinator company from the winter 2025 batch, runs the post-training: reinforcement fine-tuning with methods like GRPO and DAPO and multi-turn tool training, without the team standing up its own RL infrastructure. The output is a task-specific model that beats the foundation model on cost and latency.

As inference bills climb, a cheaper specialized model is a direct line to margin, which is the demand Osmosis is chasing. The CEO sold a gaming startup before this; the CTO led real-time recommendations at TikTok. The threat is plain: the big labs are folding reinforcement fine-tuning into their own platforms.