The raise is not the state
Doc ATL-109 · Rev 2026-08-19
Contents
I keep a private wire. It is not a newsletter and it is not my voice — it is other people’s feeds, ingested, so I can read the tape without opening twenty tabs. The last fourteen days printed a lot of “raises.” Most of what printed is noise: a fund closing, a CEO on a Bloomberg hit, a legal AI “in talks” at a number somebody leaked. Below is the closed tape, then the rumor column. Valuations are not state. Cash in the bank is closer.
This is a reading, not original reporting. Every dollar is cited to the outlet that printed it. Nothing here is investment advice.
What closed#
Amounts and marks as printed. “Lead” is whoever the outlet named first. Company names go to the company; the source column is the article.
| Date | Company | What they are | Amount | Mark | Lead / named | Source |
|---|---|---|---|---|---|---|
| 18 Aug | Etched | inference clusters | $700M | $21B | Jane Street | TechCrunch |
| 17 Aug | Higgsfield | AI video | $400M B | $5.4B | DST Global | TechCrunch |
| 17 Aug | Groq | Nvidia neocloud | $350M | $3.5B | Disruptive; Nvidia planned | TechCrunch |
| 17 Aug | Wispr | dictation → meetings | $280M B | $2B | Menlo | TechCrunch |
| 17 Aug | Gravis Robotics | construction robots | $200M | $1B | SoftBank | Sifted |
| 17 Aug | Terra Industries | defense, Global South | $52M seed | — | — | TechCrunch |
| 13 Aug | Databricks | data + agents | $5B | $190B | Coatue, Blackstone, MGX, T. Rowe, Sixth Street | TechCrunch |
| 12 Aug | Thrive Holdings | PE + OpenAI embeds | $2B | $12B | SoftBank, D1, Altimeter | TechCrunch |
| 12 Aug | Lovable | AI coding | $400M | $13.3B | EU Scaleup Fund among others | Bloomberg / Sifted |
| 12 Aug | ClearJet | cargo | $25M | — | — | Crunchbase |
| 11 Aug | Neros | defense | — | $2.5B (tripled) | — | Bloomberg |
| 10 Aug | Cambridge Aerospace | UK air defence | — | $3.4B | — | Sifted |
| 5 Aug | Moove | robotaxi ops | $250M | — | — | TechCrunch |
| 4 Aug | HappyRobot | voice agents | — | $1.2B | a16z | Sifted |
Work, innovation, moat#
What they sell, what they changed, what might keep a competitor out. Company-claimed numbers stay labeled. Closed names only.
Etched#
Etched sells frontier inference clusters, not a chip on a slide. The product is a rack: Sohu, a transformer-only ASIC, co-designed with the box, the memory, and the software that feeds it. The innovation is the refusal: they burned attention into silicon and threw away GPU generality, which is how they claim order-of-magnitude tokens-per-watt against Nvidia on text, image, and video transformers (TechCrunch, their own post). The moat, if it exists, is a customer who already installed a rack — Jane Street — plus TSMC-era supply. The risk is the same as the bet: if the world stops being transformers, Sohu is a brick.
Higgsfield#
Higgsfield is a consumer video factory: one subscription that runs other people’s models (Seedance, Veo, Kling, and a pile more) and adds the controls a raw model does not have — Soul ID for a face that survives across shots, Cinema Studio for camera and lens, a commercial pipeline. They are not a foundation-model lab. The innovation is the control layer and the claim that you can sit there and make a feature (Hell Grind is their proof-of-life). The moat they advertise is 30M users and $700M annualized revenue. I am writing that as a claim. Underneath, the engines are rented. Distribution and habit are the only things that do not walk.
Groq#
Groq used to make an LPU — a deterministic inference chip. In December Nvidia licensed the stack and hired Jonathan Ross. What raised this week is the leftover: a neocloud. They rent Nvidia boxes across 13 sites, 54 MW going to 200 MW (TechCrunch). The innovation walked with the founder. The current company sells capacity and software around someone else’s silicon. The moat is sites, sovereign relationships, and whatever LPU residue Nvidia did not take. It is not an ASIC story anymore.
Wispr#
Wispr makes Flow, a voice-to-text layer that types into every app on Mac, Windows, iPhone, and Android. This round pays for the expansion: Notetaker, meetings without a bot in the call, and Canto, their own model, after users said quality dipped (TechCrunch). The innovation is OS-level insertion plus a custom dictionary that syncs across devices — you train it once, you speak everywhere. The moat is the habit and that dictionary. Meetings is where the money is: the transcript becomes the object agents read.
Gravis Robotics#
Gravis does not sell a new excavator. ETH Zurich spin-out, 2022. The product is the Gravis Rack: a bolt-on kit plus an OS that runs mixed fleets — Caterpillar, Case, John Deere, JCB, Hitachi, Volvo, the rest (their Series A note). The innovation is retrofit. You do not wait for OEM autonomy; you put a rack on the machine that is already on site. The moat is installed-base across brands and a UK CAM Pathfinder with Flannery, the country’s largest plant-hire. SoftBank is buying the labor-shortage story in a trillion-dollar industry that still runs on a human in a cab.
Terra Industries#
Terra is trying to be the defense prime for Africa and the Global South: Archer VTOL, Iroko UAV, Kama interceptor, Duma UGV, sentry towers. Founded 2024 in Nigeria; Pax-2 in Accra is pitched as the continent’s largest drone factory, tens of thousands of airframes a year by 2028 (their site, TechCrunch). The innovation is local manufacture for African militaries, mines, and pipelines instead of importing an Anduril kit. The moat, if the factory and the contracts are real, is geography and offtake. A seed at this ambition is a factory bet, not a software one.
Databricks#
Databricks is the enterprise lakehouse. What they added this cycle is Lakebase: serverless Postgres for agent state, sitting on the same platform as the warehouse, after they bought Neon. The innovation is collapsing OLTP, OLAP, and governance so an agent can remember, write, and be audited without a second database. They claim $7B ARR, +80%, cash-flow positive, Lakebase at $100M run-rate, a three-cloud commit (TechCrunch). The moat is switching cost: Unity Catalog, the existing spend, and the fact that $15B of demand showed up when they did not need the money.
Thrive Holdings#
Thrive Holdings does not sell a model. It buys accounting firms and IT shops, then OpenAI seconds research and engineering into them in exchange for equity (OpenAI’s own note, TechCrunch). The innovation is the embed: the lab sits inside the workflow company, not in a slide about “AI transformation.” The third vertical is permits and physical plant. The moat is ownership of the target firms plus OpenAI labor that a competitor cannot hire on a statement of work. Implementation is the product.
Lovable#
Lovable is vibe coding for people who do not write code: you describe an app, it emits a full-stack Next/React/Supabase thing with auth, a database, and a deploy button. Agent Mode then owns multi-file changes. The innovation is end-to-end generation rather than a snippet in an IDE — that is the difference with Cursor. The moat is distribution among non-technical founders. The code is not a moat: Bolt, v0, Replit do the same job, and Lovable will GitHub-export the repo, which is honest and also means they do not lock you in.
ClearJet#
ClearJet is the Uber of belly cargo. They do not own planes. They buy unused capacity on commercial flights already flying between US cities, stuff e-commerce parcels into custom containers, and let an AI pick the path on cost, speed, and geography (Crunchbase). The innovation is the license plus the container plus the router — an indirect air carrier sitting on 95 airports. The moat is airline relationships and zip coverage, not metal. Asset-light until an airline decides it wants the margin back.
Neros#
Neros makes US-made attritable drones. Flagship is Archer, a BlueUAS FPV; the new money is for Archer AI (terminal guidance, GPS-denied hold) and Bandit, a kinetic interceptor aimed at Shahed-class threats, plus a Torrance factory pointed at a million airframes a year by 2028 (their C note, Bloomberg). The innovation is not a novel airframe. It is a compliant supply chain and a production line. The moat is offtake from the Army and Marines plus the factory. If China still makes the internals, the story is a press release.
Cambridge Aerospace#
Cambridge Aerospace builds cheap interceptors. Skyhammer: Mach 0.7, ~30 km, tens of thousands of dollars, meant to kill Shaheds instead of firing a Patriot at a $20k drone. Starhammer: Mach 2, shorter range, cruise-missile class. They are vertically integrating rocket motors in Norfolk (Sifted, UK MoD tests in Jordan). The innovation is the cost curve. The moat is a British government contract plus a motor plant. Fourteen months from sketch to contract is the other sentence.
Moove#
Moove started as income-based vehicle finance for African ride-hail drivers. The company that raised this week is the ops layer under Waymo: charging, depots they call Nests, inspection, repair, in Phoenix, Miami, Las Vegas, London next. They still run ~42,000 human-driven cars (TechCrunch). The innovation is admitting that autonomy software is not the bottleneck — the depot is. The moat is the Waymo contract and the muscle memory of running a fleet that already has to make money every day.
HappyRobot#
HappyRobot builds voice agents that do freight work: check calls, carrier sales, track-and-trace, talking into a TMS, not just reading a script. They train their own models on logistics calls; a16z wrote the first check on that. DHL, Kuehne+Nagel, Ryder are named. They are expanding the same worker into utilities and airline cargo (Sifted). The innovation is the write-back, not the voice. The moat is the workflow install. A better voice model does not move a broker who already wired HappyRobot to the load board.
Grounded#
Grounded upfits modular interiors onto van chassis in Detroit — living space, clinic, command post, coffee shop. They started on Ford E-Transit / BrightDrop; when those EV platforms died they started doing gas vans too (TechCrunch). The innovation is the cube interior plus Grounded+ electrics, not the vehicle. The moat is thin: speed of a custom van in days, a factory, Tesla/SpaceX alumni on the floor. Anyone with a jig can copy the furniture. The software in the wall is the only piece that does not unscrew.
Wordsmith#
Wordsmith is an operating layer for in-house legal: intake, contract review, vendor analysis, HR drafting, now pushing into financial services and insurance. Edinburgh, agents not a chatbot (Sifted, their B-extension note). The innovation is putting the agent on the department’s actual queue instead of a Copilot sidebar. The moat is named logos (BT and 500 others, claimed) and the files already in the system. Legal AI is crowded. Whoever sits on the intake wins.
Moss#
Moss is European spend management: cards, invoices, reimbursements, real-time budgets, 5,000+ companies, claimed €6.5B of spend a year and €70M ARR (Sifted). This round is the pivot to Finance AI — agents on coding, close, AR, the jobs a controller actually hates. The innovation was bundling the card with the ledger so the agent has something to read. The moat is the existing spend graph in DACH and the UK. A new agent without the card is a demo.
The LP tape, separately, because a fund close is not a company: Reach Capital, $265M Fund V (TechCrunch). QuantumLight (Storonsky), $500M fund II (Sifted). Accel, $800M Europe early-stage plus a $3.5B global AI vehicle (Sifted / Bloomberg).
Four sentences that are actually the week#
1. Jane Street bought a rack, then led the round. Etched’s mark went $5B (Dec) → $10.3B (23 Jul) → $21B (18 Aug). The blog post that matters is not the valuation. It is Jane Street saying they tested the chip, liked the early results, and have a rack in their own datacenter. That is a customer who can write a check the size of a Series C. Everything else is a multiple.
2. Groq is no longer the company on the hoodie. After Nvidia hired Jonathan Ross and licensed the stack, Groq is raising as a neocloud — Nvidia boxes, 13 sites, 54 MW → 200 MW. $3.5B versus $6.9B last September. They say it is not a down round because it is a different firm. Fine. Then stop reading it as an ASIC story.
3. Databricks did not need the money. The phone book did. Ghodsi wanted $1B. The Information printed a fundraise in the middle of their conference. $15B of demand showed up. They issued more stock. Coatue and a paragraph of late-stage names. Claimed state: $7B revenue run-rate, +80%, cash-flow positive; Lakebase $100M run-rate; three-cloud commit; a 100-person research bench; M&A (Electric / PGlite this week). This is what a private mega-cap looks like when the public calendar is optional.
4. Implementation is now a product. Thrive Holdings — Thrive Capital spinout, OpenAI on the cap table, engineers seconded into acquired firms — took $2B at $12B to buy more accounting and IT shops and stand up a third vertical around permits and physical plant. SoftBank, D1, Altimeter. If you thought the next trillion of AI was another model, the late-stage book disagrees. It is embedding.
Higgsfield is the consumer-compute cousin: $400M B, $5.4B, eight months after $1.3B, DST-led. Company-claimed $700M annualized revenue and 30M users. I am writing that as a claim. Video is a compute bill with a UI.
Wispr is the interface cousin: $280M at $2B, Menlo, dictation into meetings, a new model (Canto) after users said quality dipped. The interface layer is getting priced like infrastructure.
The rumor column#
Not closed. Do not file these as rounds.
- Legora, legal AI, “in talks” above $10B (Sifted / FT).
- Tekever, surveillance drones, talks at €5.5B (Bloomberg).
- Uforce, drones, “targeting” $4B (Sifted).
- A Palmer Luckey–backed startup bank, “set to raise” $1.5B (FT).
- Intel share sale “said to” $20B (Bloomberg) — public company, different animal.
- Kalshi at $40B is a talking-head mark on a Bloomberg hit, not a close.
How to read it if you build agents#
The capital is not “in AI.” It is in three stacks that sit under agents:
- Tokens out the door — Etched (specialized inference), Groq (rented Nvidia), Higgsfield (video as a 60,000-word bill per minute). If your product is an agent that talks or draws, you are downstream of these invoices.
- The harness around the model — Databricks (warehouse + agent database + M&A), Lovable / Wispr (coding and voice as surfaces), HappyRobot (voice in production). This is where I live. The model is a commodity input; the loop is the company.
- Bodies and permits — Gravis, Terra, Neros, Cambridge Aerospace, Moove, Thrive’s new physical-asset vertical. Embodiment and industrial policy are no longer a side aisle.
The ads-marketplace point from last week still holds. These raises are the other side of that trade: if default rank is no longer something a human stares at, the money moves to whoever owns compute, the runtime, or the permit. Sponsored slots are a 2015 product. A Jane Street rack is a 2026 one.
I will keep the desk. Closed rounds go in the table. “In talks” stays in the rumor column. If a number only exists on a podcast, it is not in the table.
— Gökhan Turhan · Numetal Labs · Gökhan Ventures
A reading of third-party coverage from my private RSS desk, 5–18 August 2026. Not an offer, not a recommendation, not my original reporting. Company-claimed revenue (Higgsfield, Databricks run-rate) is labeled as such.