❯ AI Inference Cloud Company Groq Raises $350 Million, Valuation Halved to $3.5 Billion
[PIVOT] Groq 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 racks Nvidia GPUs in 13 data centers worldwide and sells inference capacity on a metered basis to developers and enterprises. The in-house LPU chip line is no longer the core business. The company says its platform has more than 6 million developers, enterprises, and AI-native companies.
[EXODUS] Last September, Groq was still valued at $6.9 billion — in less than a year, that figure has been cut nearly in half. What happened in between is rare in chip circles: in December 2025, Nvidia reportedly paid $20 billion for a license to Groq’s core inference technology, while simultaneously hiring away founder and CEO Jonathan Ross, President Sunny Madra, and roughly 90% of the engineering team — with the money going to existing shareholders. The corporate shell stayed; the people and the technology left. In June of this year, Groq first raised $650 million to launch its transformation, then took in this $350 million two months later. Groq itself does not concede this is a down round, saying it is repricing “the Groq that exists after the Nvidia licensing deal.” Now sitting in the CEO seat is Alex Davis, chairman of Disruptive.
[REMAINS] Strip away the chips, and what Groq truly has left is three things: 13 data centers spread across North America, Europe, the Middle East, and Asia-Pacific; current installed capacity of 54 megawatts, against a company target of over 200 megawatts for 2027; and a distribution base accumulated from more than six million developers. That hand dictates that it can only play toward a new breed of cloud, where the rivals are companies like CoreWeave that specialize in renting out AI compute — no longer Nvidia. The problem is that the financial model for this business is inherently ugly: capital expenditure is extremely heavy and typically propped up by debt, the GPUs it buys are depreciating fast, and no matter how quickly revenue grows, it may not convert into free cash flow. CoreWeave has already demonstrated this problem. Groq’s financials are still not public — outsiders cannot see its unit economics.
[RESET] In the same week, Groq’s valuation was halved for selling its chip technology to Nvidia, while Etched’s doubled for getting its racks into customers’ data centers. The two events are two sides of the same rule: the value of inference hardware lies not on the design blueprint, but in machines that are powered on and running. This $3.5 billion round is no longer buying a chip company; it is buying a compute sublessor doing business on someone else’s hardware, and the valuation framework has switched accordingly — from semiconductors to infrastructure. The expectations truly being rewritten are those of 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. Existing shareholders got their money back through the $20 billion licensing fee; the company that remained was repriced. This is an exit path more people will likely factor into their models going forward.
▪ SIGNALNvidia paid $20 billion for the technology and the people, leaving behind a shell that switched to renting GPUs. The optimal path for challengers is shifting from mass production and IPOs to being licensed and absorbed by a giant.
❯ AI chip company Etched closes $700 million Series D, valuation doubles to $21 billion
[ORDERS] Etched has closed a $700 million Series D at a post-money valuation of $21 billion, led by quant trading firm Jane Street. The company sells its in-house Sohu inference chip as a complete rack system — cabinet, software, and liquid cooling included — so customers can run large-model inference the moment it arrives in their own data center. Founded in 2022 by three Harvard dropouts, Etched has raised $1.9 billion in total and holds more than $1 billion in signed customer contracts.
[ASCENT] Last December the company was valued at $5 billion; in July it raised $300 million at a $10.3 billion valuation; one month later it jumped to $21 billion — more than quadrupling in eight months. The inflection point is clear. Etched only came out of stealth at the end of June, with a public scorecard of $800 million raised and $1 billion in orders — but zero machines delivered. What actually changed the valuation curve was last month’s delivery: the first rack went into Jane Street’s data center, and after testing, Jane Street said the results were “satisfactory” and immediately routed live production workloads through it. A customer tested the machine and came back to lead the round — about as hard an endorsement as a chip startup can get. The money is not betting on a roadmap; it is betting on a machine already doing work in someone else’s data center.
[THE BET] Etched’s bet rests on two pillars. One is low-voltage inference (LVI): it pushes the operating voltage of its compute units below half that of competing AI chips, buying several times the compute density per unit area; the company claims sparse trillion-parameter models can run at more than 80% of peak throughput. The other is cluster-scale memory (CSM): it mixes HBM and SRAM and adds a proprietary low-latency interconnect so every chip in the rack shares a single memory pool, built specifically for long contexts and multi-trillion-parameter mixture-of-experts models. Its original Sohu chip hardwires Transformer computation directly into silicon; the company cites 500,000 tokens per second on Llama 70B, roughly 20 times an eight-GPU H100 server. 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 pivoted to cloud, Cerebras mostly sells chips and cloud services, and Etched sells complete racks from day one — it will not let customers assemble their own systems. For Nvidia, the problem is that these specialized machines only compete for orders in the inference segment.
[CAPITAL'S BID] A $21 billion valuation against $1 billion in contracts on hand is a contract multiple of just over 20x — an extremely aggressive tier for a hardware company. But the pricing logic of this round is straightforward: investors are paying for inference capacity already installed, not for a chip roadmap. Over the past two years, most inference-side challengers fell at the same hurdle — great paper performance, but customers refused to move production workloads over. Etched cleared that hurdle with a single rack inside Jane Street’s data center. The real pressure now falls on inference-chip companies still at the sample and white-paper stage: the reference point has shifted from benchmarks to installed base, and the fundraising bar has moved up with it. What to watch next is delivery cadence: whether $1 billion in contracts becomes revenue depends on TSMC capacity and rack yield — not on publishing another round of benchmark scores.
▪ SIGNALA chip company’s costliest valuation jump came from a customer plugging the machine in, not from another round of benchmarks. 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 $1 billion
[WATT ANGLE] Velaura AI has closed a $110 million Series A at a post-money valuation exceeding $1 billion, led by Seligman Ventures. It doesn’t manufacture entire chips; it sells digital chip IP and a companion design platform. Customers integrate its modules into their own AI accelerators and get lower power consumption at the same compute. Titan Core, the flagship product, is positioned against conventional accelerator designs, with the company claiming 2x to 4x performance per watt.
[MODULE TO PLATFORM] This is a Silicon Valley company — its founding date is not disclosed in public materials — and it previously existed mainly as an IP supplier. The change is coming from the demand side: AI data centers’ bottleneck has shifted from compute to power. Per-rack power keeps climbing, electricity bills and cooling now determine deployment scale, and energy efficiency has turned from an engineering metric into a procurement metric. Velaura only launched Titan Core as a standalone platform this year, with the funding round following close behind. The investor roster shows who it wants to sell to: in addition to the lead, new entrants 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 miner MARA represent the advanced-process and power-intensive-deployment ends, respectively.
[TAPE-OUT RECORD] The hardest gate in the IP business is whether customers believe you can mass-produce. Velaura’s credential is more than 30 million ASICs: its underlying technology has already shipped at commercial scale across multiple advanced process nodes, with real yield and reliability track records — not simulation data. That sets it apart from most low-power startups that only have lab results. The team hails from Apple, Nvidia, Google, Qualcomm and Marvell; co-founder and CEO Rajiv Khemani has spent decades in the semiconductor industry. It is betting on two markets at once: hyperscale data centers on one side, and so-called physical AI on the other — robots, drones and autonomous systems — where power sensitivity is even greater than in data centers because the battery is only so big. Proceeds from this round will go toward accelerating product commercialization, expanding engineering and customer teams, and joint development with strategic customers.
[POWER DECIDES] A $1 billion valuation for an IP company that hasn’t yet delivered at scale — the pricing logic is that energy efficiency is becoming the hardest constraint in AI infrastructure. Capital is now paying separately for per-watt performance, something that would not have been a standalone line item two years ago. The beneficiaries are upstream suppliers that can push the power curve down; the ones under pressure are compute operators scaling by stacking cards and electricity. When power supply cannot keep up, whoever’s chip is more efficient can pack more compute into the same rack. The capacity metric for data centers has to change — previously counted by GPU count, henceforth counted by available power.
▪ SIGNALThe ceiling on compute expansion is shifting from chip supply to power supply. Energy efficiency has gone from an engineering metric to a line-item price on procurement lists.
❯ AI Video Company Higgsfield Raises $400 Million Series B at $5.4 Billion Valuation
[REVENUE FIRST] AI video and image generation platform Higgsfield has closed a $400 million Series B at a $5.4 billion valuation, led by growth fund DST Global. Users type a one-line idea into its web workspace, and the platform directly produces finished videos and images: Cinema Studio, aimed at directors, handles storyboards and multi-scene final output, while Marketing Studio, aimed at marketing teams, batches out ad creatives. Founded in San Francisco in 2023, the company has raised more than $450 million to date, with this single round accounting for most of it.
[4X IN 8 MONTHS] This January, Higgsfield’s Series A raised just $50 million at a $1.3 billion valuation; eight months later, the valuation sits at $5.4 billion — more than a quadrupling. Two things happened in between. First, Supercomputer launched in May, an agentic product that runs multi-scene visual production end to end; within three months of launch, users on that product line grew 42x. Second, the enterprise door cracked open: advertising, film/TV, fashion retail, finance and even pharma started paying. The company reports annualized revenue of $700 million — higher than the size of this round. At least 18 institutions participated; Accel, Menlo Ventures and other existing backers all added to their stakes, with Goldman Sachs Alternative Investments’ growth equity arm, Intel Capital and NTT DOCOMO Ventures also on the list.
[WHY IT WINS] Higgsfield isn’t the one with the strongest models — it was the first to turn generative tools into a production workflow. Creative tool Runway and enterprise talking-head video maker Synthesia each hold a segment; Higgsfield simply split its product lines by job function: directors use Cinema Studio to make films, marketing teams use Marketing Studio to ship ad creatives in bulk, and both lines share the same generation backend. What enterprises are buying is schedulable capacity. It has also pulled ahead on scale: more than 30 million users across 238 countries and regions, over 20 million generations per month, with 390 of the Fortune 500 using it. Its ability to close enterprise deals comes down to the team’s track record: CEO Alex Mashrabov’s previous company, AI Factory, was acquired by Snap in 2019, where he scaled consumer-grade generative visual products; CTO Yerzat Dulat oversees technology. The round’s proceeds are earmarked for R&D, global infrastructure, AI talent and overseas markets, with compute the biggest line item: video is the most compute-hungry segment of AI, and with 20 million generations a month, if costs can’t be pushed down, the gross margin on $700 million in revenue won’t hold.
[BUDGET SHIFT] A $5.4 billion valuation against $700 million in annualized revenue — under 8x revenue — counts as cheap among today’s AI companies, provided that $700 million sticks. This round, capital is buying retention, not generation quality: base-model capability gets leveled month by month, anyone can plug into comparable models, and the layer that keeps collecting money is the one that locks workflows into enterprise processes. The substance of those 390 Fortune 500 customers is ad-creative production moved in from outside agencies. That is also the most fragile spot: Higgsfield doesn’t train its own base models, and once upstream model vendors move into marketing-asset tools, this middle layer’s bargaining power erodes first. Ad production firms need to redo their math — budgets are leaving the timesheet-priced invoice.
▪ SIGNALA company that trains no models has reached $700 million in annualized revenue by slotting generation into the daily schedules of marketing and film/TV. Visual-production money is shifting from capacity providers to workflow contractors.
❯ AI Finance Software Company Rillet Raises $100M Series C at $1B Valuation
[LEDGER REWRITE] AI-native ERP company Rillet has closed a $100 million Series C at a $1 billion valuation, led by ICONIQ. It rewrites the corporate general ledger: transaction data streams into the books in real time through native integrations, and AI agents handle reconciliation and journal entries directly inside the ledger. 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.
[14 MONTHS, 3 ROUNDS] This is Rillet’s third round of funding in 14 months, bringing cumulative funding to more than $200 million. The pace is this dense because demand is moving faster than product iteration — new annual recurring revenue doubled in the past three months. It is replacing the legacy systems Oracle Fusion, SAP, Workday, Microsoft Great Plains, and NetSuite, along with the sprawl of spreadsheets and plug-ins that grew up around them. Most of those systems are based on architectures from 10 to 20 years ago: data sits in one place, work happens in another, and the two sides are reconciled only at month-end close. AI agents cannot get into that structure; they can only attach to the outside as assistants. Rillet’s approach is to replace the ledger itself so the agents have a place to stand. Disclosed customers include Mercor, Function Health, and Temporal.
[WHO'S OUT] The pitch lands on the close. Traditional finance teams spend one to two weeks each month on monthly close; Rillet argues for continuous close — data streams into the ledger in real time, current-period numbers are always visible, and the concentrated month-end cycle is flattened. Founder and CEO Nicolas Kopp puts it this way: “Finance agents don’t just need access to the data; they have to work inside the general ledger.” He predicts that within two to 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. Once the general ledger moves over, audit, tax, and consolidated reporting all follow, making renewal almost the default. That also explains why Sequoia, a16z, Bain Capital Ventures, Oak HC/FT, Battery Ventures, FirstMark, Scale Venture Partners, and Creandum have invested in three consecutive rounds within a year — the bet is on the window to capture installed base, since the ERP layer only turns over once per generation.
[GENERATION SHIFT] A $1 billion valuation is not expensive in the ERP category — NetSuite was acquired by Oracle for $9.3 billion back then. Capital is pricing whether AI can enter core systems. Over the past two years, enterprise AI money has mostly gone to edge use cases — customer service, documents, code completion — where the impact of incidents is containable. The general ledger is another matter: one wrong entry leaves a trace on the audit report. Rillet’s $1 billion valuation means the market has accepted that AI agents can execute operations inside an audited system, as long as approvals and audit trails are in place. Under pressure are the companies that attach plug-ins to the outside of legacy ERP; their position rests on the premise that old systems cannot be replaced, and that premise is loosening.
▪ SIGNALEnterprise AI money is moving from edge use cases into audited core systems. Replacing the general ledger layer locks in a decade.
❯ Voice Input Company Wispr Raises $280M Series B at $2B Valuation
[SPEAK TO TYPE] Voice input company Wispr has completed 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 field and Flow converts spoken language directly into polished written text. The company has raised $361 million in cumulative funding, and more than 60 billion words have been written on the platform.
[FAST FOLLOW-ON] The previous round was a $25 million Series A extension led by Notable Capital in November 2025, bringing cumulative funding to $81 million at the time. In less than ten months, that figure has jumped to $361 million. In between, it has moved itself up a layer from a dictation tool. Enterprise adoption has been fast; the company says nearly every Fortune 500 company and more than 10,000 businesses have people using it. These tools typically enter through employees installing them themselves, and then IT procurement takes over once they become a habit. Existing investors Notable Capital, NEA, Neo Ventures, 8VC, and MVP Ventures all added to the round, joined by new backers Acrew, Forerunner, Goodwater, and Peak XV, with a string of athletes and celebrities on the list.
[IN-HOUSE MODEL] Wispr also previewed Canto, its first self-developed speech recognition model — the dividing line between calling on others’ models and building its own. The company’s figures: in noisy, real-world environments, the word error rate drops from above 30% to 5% to 10%, a roughly fourfold improvement; the model supports multiple languages and mid-sentence language switching, and also pulls from users’ own vocabularies and contact lists. The metric it uses to measure results is the zero-edit rate — the share of a dictation session that needs no correction at all. Wispr expects Canto to cut the number of edits needed in daily use by roughly 30%. The product line is also expanding: meeting notes tool Flow Notetaker is already live, and it has established the Wispr Advanced Interfaces Lab, led by former Amazon Alexa researcher Ariya Rastrow, with the goal of having systems understand intent and return results directly. This round’s capital will go mainly into model R&D and expanding coverage.
[INPUT LAYER] A $2 billion valuation for a dictation tool looks absurd on its face; capital is buying the position at the input layer. The keyboard is the entry point to all software; whoever stands between the input field and the application can see what users want to do in every context — which is also the confidence behind Wispr’s push toward understanding intent and delivering results directly. The risks are just as obvious. Operating-system vendors already do dictation themselves; Apple and Google could build this capability into the OS at any time. Wispr’s room to maneuver lies in cross-platform coverage and enterprise-side manageability. Voice’s positioning needs to be reassessed — it is moving from accessibility features and in-car scenarios into everyday desk input.
▪ SIGNALThis round is not a bet on transcription accuracy; it’s a bet on the input-field position. Whoever catches spoken language first sees user intent first.
❯ Satellite constellation company Muon Space closes $250M Series C, led by Eclipse
[LEAD] Satellite company Muon Space has closed a $250 million Series C round, led by Eclipse Capital, with the round oversubscribed. It doesn’t sell individual satellites — when a customer places an order, they’re buying an entire functional constellation, with mission design, satellite platforms, payloads, software, and on-orbit operations delivered as a package. The company calls this model Mission Foundry. Founded in 2021 in San Jose, California, the company’s announcement shows cumulative equity funding has now surpassed $386 million.
[SCALE] The previous rounds were the 2024 Series B and a 2025 Series B extension, and over the past year-plus Muon’s main job has been proving it can manufacture at scale. It launched 7 satellites in the first half of this year, bringing its cumulative on-orbit total to 11 satellites across six launches, with a 100% mission success rate. More than 50 satellites are currently in development, 13 of which are already on the launch manifest for the coming 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 its previous capacity. Two programs are already running: FireSat, a global wildfire monitoring constellation built with Earth Fire Alliance and Google.org, selected for Time’s 2025 Best Inventions list; and Vindlér 2.0, a radio frequency data and analytics constellation built for Sierra Nevada Corporation.
[MODEL] The space industry has traditionally been a one-off custom business — every customer’s constellation requires years of design from scratch. Muon has turned that into a replicable platform, pulling simulation, design, manufacturing, launch coordination, and on-orbit operations under one roof, compressing delivery timelines from years to months. CEO Jonny Dyer’s own words: “Space infrastructure needs to scale like cloud infrastructure.” Customers span defense, government, and commercial — this dual-use civil-military structure is what closes deals fastest right now, with stable government budgets on one side and commercial growth on the other. The investor roster tells the same story: beyond lead investor Eclipse, Google, Salesforce Ventures, Wellington Management, I Squared Capital, and Toyota-backed Woven Capital are all in — spanning cloud and software, infrastructure funds, and an automaker. One line item in where the money goes deserves special attention: on-orbit AI compute. The more data a constellation collects, the less feasible it becomes to downlink all of it to Earth; putting inference on the satellites is moving from concept to engineering schedule.
[IMPACT] Strictly speaking, this isn’t AI funding — but it locks onto the same question as the other rounds in today’s briefing: wherever data is generated, compute must follow. Capital is paying for the data sources themselves. Wildfire, RF, Earth observation — these are physical-world data that model training and real-time inference can’t obtain anywhere else, and satellites are the only collection point. The beneficiaries are integrated companies that can build, launch, and operate satellites simultaneously; the ones under pressure are component suppliers that only serve one segment. Once whole-package delivery becomes the norm, 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.
▪ SIGNALSpace is shifting from custom engineering to batch manufacturing. Whoever can deliver constellations in batches, like server cabinets, earns the valuation of an infrastructure company.
❯ Smart Ring Maker Happy Health Raises $75 Million to Target Sleep Apnea
[RING DIAGNOSIS] According to a company announcement, Happy Health has closed a $75 million Series A round led by ARCH Venture Partners and OpenLoop. It makes a clinical-grade smart ring, the Happy Ring, worn overnight to collect physiological data; AI interprets the data and issues a diagnostic conclusion for obstructive sleep apnea. The company was founded in 2019 in Austin, Texas.
[TWO CLEARANCES] Both lead investors have backed the company since its founding in 2019. The product has received two FDA clearances — routine biometric monitoring in September 2024, and at-home sleep testing in June 2025 — the latter making the Happy Ring the first smart ring approved for multi-night sleep testing. The nine months in between marked its step from consumer wearable to medical device.
[THREE NIGHTS] Traditionally, diagnosing sleep apnea requires an overnight stay in a sleep center, wired up with leads, with scheduling that can stretch for weeks. The company says wearing the ring for three nights yields a result with 98% accuracy, while also identifying other sleep problems 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 built the company around his own medical history. That is also what separates it from consumer smart rings: the latter offer trend references, while Happy delivers a diagnosis that can enter the medical record. The round’s proceeds will go toward advancing clinical validation and expanding the platform from sleep to at-home monitoring of other chronic conditions.
[HARDWARE-MEDICINE MERGE] Capital is funding AI that can make it onto insurance bills. Consumer wearables have amassed a trove of physiological data over the years, but without diagnostic qualification they count only as lifestyle products, with hardware margins as their revenue ceiling. Once FDA clearance is in hand, the payer switches from consumer to payor, and the unit economics of the same ring change entirely. The sleep-center business needs to be revalued — a diagnostic step requiring dedicated facilities and technicians is being replaced by a ring worn to sleep.
▪ SIGNALThe watershed for wearables is whether they can bill. The moment diagnostic qualification is secured, the payer shifts from consumer to insurance company.
OUTLOOK
[BILLS ONLY] Eight deals totaling $2.265 billion: three bets on silicon and compute, three on application-layer revenue, and two landed outside AI — satellites and medical devices. It looks scattered, but the basis for judgment is highly consistent: capital’s premium this cycle goes only to bills already sent. 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 credentials. Groq, unable to produce such credentials, saw its valuation cut in half. The exceptions are Velaura and Muon, both selling capacity that has yet to materialize. The beneficiaries are mid-tier players with paying customers in hand; those under pressure are peers left with only demos and white papers. The next money will still chase installed base and renewal counts. The number to watch next is capacity fulfillment rate, not how fast valuations rise.