❯ AI Inference-Cloud Company Groq Raises $350 Million, Valuation Halved to $3.5 Billion

PIVOTGroq has closed a $350 million funding round at a post-money valuation of $3.5 billion, led by investment firm Disruptive, with Nvidia planning to follow on. What it sells now is compute rental: it stands up Nvidia GPUs across 13 data centers worldwide and sells inference capacity to developers and enterprises on a metered basis. The in-house LPU chip line is no longer the main business. The company says its platform has more than 6 million developers, enterprises, and AI-native companies.

THE GUTTINGLast September, Groq was still valued at $6.9 billion — less than a year later, that has been cut by nearly half. What happened in between is rare by chip-industry standards: in December 2025, Nvidia reportedly paid $20 billion to license Groq’s core inference technology, while also hiring away founder and CEO Jonathan Ross, president Sunny Madra, and roughly 90 percent of the engineering team — the money went to legacy shareholders. The corporate entity stayed; the people and the technology left. This June, Groq first raised $650 million to kick off its transformation, then took in this $350 million two months later. Groq itself rejects the down-round label, saying this is repricing “the Groq that exists after the Nvidia licensing deal.” Now sitting in the CEO seat is Alex Davis, chairman of Disruptive.

WHAT'S LEFTStrip away the chips, and Groq truly has three things left: 13 data centers across North America, Europe, the Middle East, and Asia-Pacific; current installed capacity of 54 megawatts, against a company target of over 200 megawatts by 2027; and a distribution base built up by more than six million developers. This hand of cards dictates that it can only play toward a new kind of cloud — with rivals becoming companies of the CoreWeave type that specialize in renting out AI compute, rather than Nvidia. The trouble is that this business’s financial model is inherently unappealing: extremely capex-heavy, typically propped up by debt, with the GPUs it buys depreciating fast; even rapid revenue growth may not translate into free cash flow. CoreWeave has already demonstrated this exact problem. Groq’s financials remain undisclosed, so outsiders cannot see its unit economics.

RESETIn the same week, Groq’s valuation was halved for selling chip technology to Nvidia, while Etched’s doubled for delivering racks into a customer’s data center. The two events are two sides of one rule: the value of inference hardware lies not in the design blueprint, but in machines that are powered on and running. The $3.5 billion round is no longer buying a chip company; it is buying a compute sublandlord doing business on someone else’s hardware, and the valuation framework has switched accordingly from semiconductors to infrastructure. The ones whose expectations are truly being rewritten are the other chip teams challenging Nvidia — the best outcome may be getting licensed and absorbed, rather than holding out until mass production and an IPO. Legacy shareholders got their money back through the $20 billion licensing fee, the company left behind was repriced, and this exit path will probably be factored into more people’s models from here on.

▮ SIGNALNvidia spent $20 billion to take the technology and the people, leaving a shell that has pivoted to renting GPUs. For challengers, the optimal path is shifting from mass production and listing to being licensed and absorbed by a giant.

❯ AI Chip Startup Etched Closes $700M Series D, Valuation Doubles to $21B

ORDERS FIRSTEtched closed a $700 million Series D at a post-money valuation of $21 billion, led by quantitative trading firm Jane Street. It sells its in-house Sohu inference chip as a complete rack — cabinet, software, and liquid cooling bundled in — so customers can run large-model inference as soon as it lands in their own data center. Founded in 2022 by three Harvard dropouts, the company has raised $1.9 billion cumulatively, with signed customer contracts now exceeding $1 billion.

8-MONTH CLIMBLast December, Etched was still valued at $5 billion. This July, when it raised $300 million, the figure was $10.3 billion. A month later it jumped to $21 billion — up more than fourfold in eight months. The inflection point is clear. Etched only emerged from stealth at the end of June, and its public scorecard at the time showed $800 million raised and $1 billion in orders — with not a single machine delivered. What actually rewrote the valuation curve was last month’s delivery: the first rack went into Jane Street’s data center. After testing, the firm called the results “satisfactory” and immediately routed its own live production load through the machine. A customer tests the box, then turns around and leads the round — in chip startups, that is about as hard a vote of confidence as exists. The money is not backing a roadmap; it is backing a machine already working in someone else’s data center.

POSITIONINGEtched’s bet sits on two pillars. First, low-voltage inference (LVI): it pushes the operating voltage of its compute units below half that of rival AI chips, buying several times the compute density per unit area — the company says sparse trillion-parameter models can sustain more than 80% of peak throughput. Second, cluster-scale memory (CSM): it mixes HBM with SRAM and adds a proprietary low-latency interconnect so chips across the whole rack share a single memory pool, purpose-built for long-context and multi-trillion-parameter mixture-of-experts models. The original Sohu chip hardwires Transformer computation directly into silicon; Etched’s published figure is 500,000 tokens per second on Llama 70B, roughly 20x a full eight-GPU H100 system. The A0 tape-out runs on TSMC’s N4P process. More than 400 engineers come from Nvidia, Google’s TPU team, Broadcom, SK hynix, and TSMC. Among inference-chip challengers, Groq has already veered into cloud, and Cerebras mostly sells chips and cloud services; Etched sells the whole rack from day one and won’t let customers assemble their own systems. For Nvidia, the trouble is that these dedicated machines only compete with it for orders in the inference segment.

WHAT CAPITAL BUYSA $21 billion valuation against $1 billion in contracts on hand — a contract multiple of just over 20x — sits in an extremely aggressive tier for a hardware company. But the pricing logic of this round is not complicated: investors are paying for inference capacity that is already installed, not for a chip roadmap. Over the past two years, most inference-side challengers died at the same spot: great numbers on paper, but customers would not move production workloads over. Etched cleared that hurdle with one rack inside Jane Street’s data center. The real pressure now falls on inference-chip companies still stuck at the sample and white-paper stage; the yardstick has switched from benchmark scores to installed base, and the fundraising bar has risen accordingly. What to watch from here is delivery cadence: whether the $1 billion in contracts becomes revenue depends on TSMC capacity and rack yield — not on issuing yet another set of benchmark scores.

▮ SIGNALA chip company’s most expensive valuation jump came from a customer plugging the machine in, not from another round of benchmark scores. Pricing power in inference hardware is shifting from performance claims to machines actually installed.

❯ AI chip IP company Velaura closes $110M Series A, valuation tops $1B

WATT ANGLEVelaura AI has closed a $110 million Series A at a post-money valuation above $1 billion, led by Seligman Ventures. It doesn’t build entire chips; it sells digital-chip IP and a companion design platform. Customers integrate its modules into their own AI accelerators to get the same compute at lower power. Its flagship product, Titan Core, targets conventional accelerator designs, and the company claims 2–4x performance per watt.

MODULE TO PLATFORMThis is a Silicon Valley company whose founding date isn’t disclosed in public materials; it previously existed mainly as an IP supplier. The change comes from the demand side: the bottleneck in AI data centers has shifted from compute to electricity. Per-rack power keeps climbing, electricity bills and cooling are beginning to determine deployment scale, and only now has energy efficiency gone from an engineering metric to a procurement criterion. Velaura only launched Titan Core as a standalone platform this year, and the funding followed right behind. The investor roster shows who it wants to sell to: beyond the lead, new backers include Capricorn Investment Group and Prosperity7 Ventures, while existing shareholders Mayfield, Maverick Silicon, MARA, Premji Invest, Samsung Catalyst Fund and StepStone Group all followed on. Samsung and crypto-mining company MARA sit at the two ends of the spectrum — advanced process nodes and power-dense deployments, respectively.

TAPE-OUT RECORDThe hardest gate in the IP business is whether customers believe you can ramp to volume production. Velaura’s proof is more than 30 million ASICs: its underlying technology has already shipped at commercial scale across multiple advanced process nodes, with a proven record on yield and reliability — not simulation data. That sets it apart from most low-power startups that only have lab results. The team comes from Apple, Nvidia, Google, Qualcomm and Marvell; co-founder and CEO Rajiv Khemani has spent decades in the semiconductor industry. It’s betting on two markets at once: hyperscale data centers on one side, and so-called physical AI — robots, drones and autonomous systems — on the other. The latter is even more power-sensitive than data centers; the battery is only so big. The round’s proceeds will go to accelerating product commercialization, expanding engineering and customer teams, and pursuing joint development with strategic customers.

POWER BILLS DECIDEA $1 billion valuation for an IP company that hasn’t yet delivered at scale rests on the view that energy efficiency is becoming the hardest constraint in AI infrastructure. Capital is starting to pay separately for performance per watt — two years ago, that wouldn’t have gotten its own line item. The beneficiaries are upstream suppliers that can bend the power curve down; the pressure falls on compute operators that scale by piling on cards and power. When electricity supply can’t keep up, whoever’s chip uses less power can pack more compute into the same rack. The capacity metric for data centers needs to change: it used to be counted in GPUs; going forward, it’s counted in available power.

▮ SIGNALThe ceiling on compute expansion is shifting from chip supply to power supply. Energy efficiency has already gone from an engineering metric to a price line on the procurement list.

❯ AI video company Higgsfield closes $400 million Series B at $5.4 billion valuation

REVENUE FIRSTAI video and image generation platform Higgsfield has closed a $400 million Series B at a $5.4 billion valuation, led by growth-stage fund DST Global. Users type a one-line idea into its web workspace and the platform produces finished-grade video and images: Cinema Studio, built for directors, handles storyboards and multi-scene final cuts; Marketing Studio, built for marketing teams, mass-produces ad creatives. Founded in San Francisco in 2023, the company has raised over $450 million in total — this round alone accounts for most of it.

4X IN 8 MONTHSIn January, Higgsfield’s Series A raised just $50 million at a $1.3 billion valuation. Eight months later it’s worth $5.4 billion, up more than fourfold. Two things happened in between. First, Supercomputer — an agentic product launched in May that auto-runs multi-scene visual production end to end — grew users on that line 42x in its first three months. Second, the enterprise door got pried open: advertising, film and TV, fashion retail, finance, even pharma started paying. The company reports annualized revenue of $700 million, higher than the round itself. At least 18 institutions joined; Accel, Menlo Ventures, and other existing backers all added capital, with growth equity from Goldman Sachs Alternatives, Intel Capital, and NTT DOCOMO Ventures also on the list.

WHY IT WINSHiggsfield isn’t the strongest model shop — it’s the first to turn generation tools into a production workflow. Runway owns creator tools, Synthesia owns corporate talking-head video; Higgsfield splits product lines straight down job functions. Directors get Cinema Studio to make films, marketing teams get Marketing Studio to produce assets, and both lines share the same generation backend. What enterprises buy is production capacity that slots into a schedule. It has also pulled ahead on scale: 30 million+ users across 238 countries and regions, 20 million+ generations per month, and 390 of the Fortune 500 on the platform. Closing enterprise deals comes down to team pedigree — CEO Alex Mashrabov’s prior company, AI Factory, was acquired by Snap in 2019, and he scaled consumer generative visual products at Snap; CTO Yerzat Dulat runs the technology. This round’s proceeds are earmarked for R&D, global infrastructure, AI talent, and overseas markets, with compute the biggest line item: video is AI’s most compute-hungry slice, and at 20 million generations a month, costs can’t be squeezed — the gross margin on $700 million in revenue won’t hold otherwise.

BUDGET SHIFTA $5.4 billion valuation on $700 million in annualized revenue is under 8x revenue — cheap for today’s AI companies, provided that $700 million sticks. This round, capital is buying retention, not generation quality: underlying model capability is flattening month over month, and since anyone can plug into comparable models, the layer that keeps collecting is the one that locks workflows into enterprise processes. The substance of those 390 Fortune 500 customers: ad creative production is moving in-house from outside agencies. That’s also the most fragile point — Higgsfield doesn’t train its own foundation models, and if upstream model vendors build marketing creative tools, this middle layer’s pricing power erodes first. Ad production firms need to redo the math — budgets are walking off the hourly quote sheet.

▮ SIGNALA company that trains no models built $700 million in annualized revenue by slotting generation into the daily schedules of marketing and film/TV work. Money in visual production is flowing from capacity suppliers to workflow contractors.

❯ AI finance software company Rillet raises $100 million Series C at $1 billion valuation

LEDGERAI-native ERP company Rillet has raised a $100 million Series C at a $1 billion valuation, led by ICONIQ. It rewrote the corporate general ledger: transaction data streams into the books in real time via native integrations, and AI agents handle reconciliation and journal entries directly in the ledger, while humans retain approval rights and every change leaves an audit trail. The company came out of stealth in 2024 and now has more than 600 customers, including public companies.

PACEThis is Rillet’s third financing round in fourteen months, with cumulative funding surpassing $200 million. The pace is that tight because demand is moving faster than product iteration: net new annual recurring revenue doubled in the past three months. It replaces legacy systems such as Oracle Fusion, SAP, Workday, Microsoft Great Plains, and NetSuite — and the spreadsheets and plugins that grew up around them. Most of these systems are based on architecture from ten to twenty years ago: data lives in one place, work happens in another, and the two sides are only reconciled at month-end close. AI agents cannot enter that structure; they can only attach to the outside as assistants. Rillet’s approach is to replace the ledger itself, giving agents a place to stand. Disclosed customers include Mercor, Function Health, and Temporal.

DISPLACEMENTThe pitch centers on closing the books. Traditional finance teams spend one to two weeks per month on month-end close. Rillet argues for continuous close: data flows into the ledger in real time, current-period numbers are available at any time, and the concentrated month-end cycle is flattened out. Founder and CEO Nicolas Kopp puts it this way: “Financial agents need more than just access to data; they need to work inside the general ledger.” He predicts that within two or three years every company will run finance this way. The moat is switching costs — ERP is one of the hardest systems in a company to replace, and once the general ledger moves over, audit, tax, and consolidated reporting all follow. Renewal is almost the default option. That also explains why Sequoia, a16z, Bain Capital Ventures, Oak HC/FT, Battery Ventures, FirstMark, Scale Venture Partners, and Creandum invested three rounds in a row within a year: the bet is on capturing the installation window, because the ERP layer gets replaced once per generation.

SHIFTA $1 billion valuation is not expensive by ERP standards — NetSuite was acquired by Oracle for $9.3 billion back then. Capital is pricing whether AI can get into core systems: over the past two years, enterprise AI spending has mostly gone to edge use cases — customer service, documents, code completion — where failures have manageable impact. The general ledger is another matter: one wrong entry leaves a mark in the audit report. Rillet getting a $1 billion valuation means the market accepts that AI agents can operate inside audited systems, as long as approvals and audit trails are properly in place. The pressure is on companies that build plug-ins on top of old ERPs — their position rests on the assumption that legacy systems cannot be replaced, and that assumption is loosening.

▮ SIGNALEnterprise AI money is moving from edge use cases into audited core systems. A general-ledger generation change locks in a decade.

❯ Voice-Input Company Wispr Closes $280M Series B at $2B Valuation

VOICE REPLACES TYPINGVoice-input company Wispr has closed a $280 million Series B at a $2 billion valuation, led by Menlo Ventures. Its product, Flow, is a cross-app dictation layer: on iOS, Android, and Windows, users speak into any input box and Flow converts the speech into clean, formatted written text on the spot. The company has raised $361 million in total, and more than 60 billion words have already been written on the platform.

10-MONTH RE-RAISEThe previous round was a $25 million Series A extension led by Notable Capital in November 2025, when cumulative funding stood at $81 million. In less than ten months, that figure has jumped to $361 million. In between, Wispr moved itself up a notch from a dictation tool. Enterprise penetration has been fast: the company says people at nearly every Fortune 500 company and more than 10,000 enterprises are using it. These tools typically get in through employees installing them on their own, with IT procurement taking over once they catch on. All existing investors — Notable Capital, NEA, Neo Ventures, 8VC, and MVP Ventures — added to the round, joined by new investors Acrew, Forerunner, Goodwater, and Peak XV, with a batch of athletes and celebrities on the list as well.

IN-HOUSE MODELAlongside the round, Wispr also previewed Canto, its first self-developed speech recognition model — the dividing line between calling someone else’s model and building in-house. The company’s figures: in noisy real-world environments, word error rate drops from above 30% to 5%–10%, roughly a fourfold improvement. The model supports multiple languages and mid-sentence language switching, and draws on the user’s own vocabulary and contact list. Its effectiveness metric is zero-edit rate, the share of a dictation session that comes out needing no edits at all. Wispr expects Canto to reduce what needs fixing in daily use by about another 30%. The product line is expanding, too: meeting transcription tool Flow Notetaker is already live, and it has set up the Wispr Advanced Interfaces Lab, led by former Amazon Alexa researcher Ariya Rastrow, with a mission to make systems understand intent and deliver results directly. Proceeds from this round are mainly directed toward model R&D and expanding coverage.

THE INPUT LAYERA $2 billion valuation for a dictation tool looks absurd on the surface, but what the capital is buying is the position at the input layer. The keyboard is the gateway to all software — whoever stands between the input box and the application sees what users want to do in every scenario. That is also what emboldens Wispr to push toward understanding intent and delivering results directly. The risks are just as obvious: OS vendors already ship their own dictation, and Apple and Google could build this capability into the system at any time. Wispr’s room to maneuver is cross-platform reach and enterprise-side manageability. Voice’s positioning needs to be reassessed — it is moving from accessibility features and in-car scenarios into everyday input at the office desk.

▮ SIGNALThis round isn’t betting on transcription accuracy — it’s betting on the position at the input box. Whoever catches the spoken word sees the user’s intent first.

❯ Satellite constellation company Muon Space closes $250M Series C led by Eclipse

TURNKEY CONSTELLATIONSSatellite company Muon Space has completed a $250 million Series C, led by Eclipse Capital, with the round oversubscribed. It doesn’t sell individual satellites — customers order an entire usable constellation, with mission design, satellite platforms, payloads, software, and in-orbit operations delivered as a package. The company calls this model Mission Foundry. Founded in 2021 in San Jose, California, the company’s disclosures show cumulative equity funding now exceeding $386 million.

FROM 11 TO 500The prior rounds were a 2024 Series B and a 2025 Series B extension. Over the past year-plus, Muon has focused on proving it can scale production. In the first half of this year it launched 7 satellites, bringing the cumulative on-orbit count to 11 across six launches, with a 100% mission success rate. More than 50 satellites are currently in development, 13 of which are already slotted into launch schedules over the next year. The real change is on the ground: a new San Jose factory just came online with a designed annual capacity of 500 satellites — ten times the previous level. Two programs are already running: FireSat, a global wildfire-monitoring constellation built with Earth Fire Alliance and Google.org, named one of TIME’s Best Inventions of 2025; and Vindlér 2.0, an RF data and analytics constellation built for Sierra Nevada Corporation.

FOUNDRY MODELThe space industry traditionally builds one-off missions — each customer’s constellation is designed from scratch over years. Muon has turned this into a replicable platform, pulling simulation, design, manufacturing, launch coordination, and in-orbit operations under one roof, compressing delivery timelines from years to months. CEO Jonny Dyer’s own words: “Space infrastructure needs to scale the way cloud infrastructure did.” Customers span defense, government, and commercial — a dual-use structure that wins contracts most smoothly right now, with stable government budgets and commercial-side growth. The investor lineup is telling as well: beyond lead investor Eclipse, Google, Salesforce Ventures, Wellington Management, I Squared Capital, and Toyota’s Woven Capital all came in — a mix of cloud and software players, infrastructure funds, and an automaker. One line in the use of funds deserves a standalone mention: on-orbit AI compute. The larger the data volumes constellations collect, the less feasible it becomes to transmit everything back to the ground; putting inference on the satellites themselves is moving from concept into engineering schedules.

VALUATION ANCHORStrictly speaking, this isn’t AI funding, but it converges with the other deals in this briefing on the same question: wherever data is generated, compute must follow. Capital is paying for the data source itself — wildfires, RF signals, Earth observation. These are physical-world datasets that model training and real-time inference cannot reach, and satellites are the only collection point. The beneficiaries are integrated companies that can build, launch, and operate satellites simultaneously; the pressure is on parts suppliers that cover only one segment — once turnkey delivery takes hold, the pricing power of the middle links gets squeezed. Capacity delivery is the only real test for this money — between the designed annual capacity of 500 satellites and the 11 currently on orbit sits a full order of magnitude.

▮ SIGNALThe space business is shifting from custom engineering to volume manufacturing. Whoever can deliver constellations in batches like server racks gets the infrastructure-company valuation.

❯ Smart-Ring Company Happy Health Raises $75M, Targeting Sleep Apnea

LEADAccording to a company announcement, Happy Health has closed a $75 million Series A round led by ARCH Venture Partners and OpenLoop. The company makes a clinical-grade smart ring, the Happy Ring, worn during sleep to collect physiological data that AI interprets to issue a direct obstructive sleep apnea diagnosis. The company was founded in Austin, Texas in 2019.

TWO CLEARANCESBoth lead investors have backed the company continuously since its 2019 founding. The product holds two FDA clearances — routine biometric monitoring in September 2024, and at-home sleep testing in June 2025 — the latter making Happy Ring the first smart ring authorized for multi-night sleep testing. The nine months between them marked its crossing from consumer wearable to medical device.

THREE NIGHTSTraditionally, diagnosing sleep apnea requires an overnight stay at a sleep center, wired up with leads, with scheduling often taking weeks. The company says three nights of wearing the ring is enough to produce a result, with 98% accuracy, and it can also identify other sleep issues such as insomnia. Founder and CEO Dustin Freckleton is a physician by training; he suffered a stroke at 24, later traced to sleep apnea, and he built the company around his own medical history. This is also the dividing line between it and consumer smart rings: the latter offer trend references, while Happy issues a diagnosis that can be entered into the medical record. The round’s proceeds will go toward advancing clinical validation and broadening the platform from sleep to at-home monitoring of other chronic conditions.

CONVERGENCECapital is paying for AI that can be billed to insurers. Consumer wearables have spent years accumulating physiological data, but without diagnostic qualification they count as lifestyle products, with hardware gross margin as the revenue ceiling; once FDA clearance is obtained, the payer switches from consumer to insurer, and the unit economics of the same ring change completely. The sleep-center business needs to be revalued — a diagnostic step that required dedicated facilities and technicians is being replaced by a ring worn to bed.

▮ SIGNALThe watershed for wearables is whether they can generate a bill. The moment diagnostic qualification is obtained, the payer shifts from consumer to insurance company.

OUTLOOK

BILLS ONLYEight deals totaling $2.265 billion — three betting on silicon and compute, three on application-layer revenue, and two outside AI in satellites and medical devices. It looks scattered, but the basis for judgment is highly consistent: capital’s premium in this round goes only to bills already issued. Etched’s $1 billion contract, Higgsfield’s $700 million annualized revenue, Rillet’s doubled new annual fees, and Happy Ring’s two clearances are all verifiable proof; Groq, lacking such proof, saw its valuation cut in half. The exceptions are Velaura and Muon, which are selling capacity not yet delivered. The beneficiaries are mid-tier players that already have paying customers; under pressure are peers left with only demos and white papers. The next money will keep chasing installed base and renewal counts. The metric to watch now is capacity realization rate, not how fast valuations rise.