Research & Intelligence

A $21M combined Seed + Series A (Costanoa Ventures lead, Emergence Capital) for AI voice agents that place and qualify mortgage sales calls — claiming to help originate $1B+ per month, $10M ARR in ~18 months, and “6 of the 10 largest independent mortgage banks” as customers. The team is credible (ex-VoiceOps/LinkedIn and ex-Google ML), and the vertical is real. But every headline number is company-sourced and unaudited, with internal inconsistencies, and the product sits on the single most litigated fact pattern in fintech: an artificial voice contacting consumers about credit. Since the FCC’s 2024 ruling that’s a TCPA violation at $500–$1,500 per call with no cap — and in Feb 2026 Lamb v. Mortgage One became a class action over exactly this. The structural tell: Sela’s site advertises “TCPA Compliant” and “Audit-Ready” while its actual privacy policy is GDPR-flavored boilerplate that names only Google Analytics and is silent on call recording, voice data, consent, GLBA, and subprocessors — and it never says where automated “qualifying” ends and licensable loan origination begins.

A $22M Series A (led by .406 Ventures, with Emerson Collective, Forerunner, Serena Ventures and Japan’s MPower Partners; ~$45M total) for the Teal Wand — a speculum-free, at-home HPV self-collection device for cervical-cancer screening, timed to a 2027 mandate requiring insurers to cover self-collected tests at no cost. The clinical foundation is real and the team is strong, but the three load-bearing claims each carry an asterisk. “FDA-authorized” is a De Novo authorization, not the PMA “approval” coverage implies; the “first and only” lost its “only” six days before the raise, when Everlywell + Waters launched a rival at-home kit; and the 2027 catalyst flows through HRSA authority now under active challenge on remand after Kennedy v. Braidwood. The risk Teal never names: it is a telehealth-and-device wrapper around Roche’s cobas assay — a supplier that already sells its own self-collection solution and can go direct — while a positive result still routes to the in-person colposcopy the product exists to avoid.

A $12M seed ($15M total incl. $3M non-dilutive; led by Temasek deep-tech builder Xora Innovation, with Foothill Ventures and SEEDS) for a Singapore firm that wants to grow single-crystal 2D-semiconductor films — MoS₂ and other transition-metal dichalcogenides — on 300mm wafers and sell the tools to foundries. The team is the strongest asset: an ex-TSMC chief scientist and Stanford’s Philip Wong advising. But the honest framing in its own materials is also the whole risk — the hard part isn’t growing the film, it’s integrating it into a fab at cost and yield parity with silicon, where contact resistance, not crystal quality, is the wall the field has stalled on for a decade. “Industry’s first” is contested by CDimension (already shipping) and China’s JiMoXin; there is no named customer, no design-win, and no revenue; and $15M is a rounding error against semiconductor-equipment economics, with commercialization self-dated to 2030–2035.

A $25M seed (led by Bain Capital Ventures, with BoxGroup and Definition Capital) for an “AI-native” health-benefits brokerage aimed at the ~6.4M U.S. businesses with 1–500 employees. The underserved market is real and the labor-leverage bet is credible — but the pitch has a conflict at its core. Corridor markets itself as “free — no fees, no markups,” yet brokers are paid commissions by carriers, a percentage of premium, so the broker earns more when premiums are higher — the exact number it promises to cut by 20%. The commission structure is never disclosed; the privacy policy never mentions HIPAA or PHI; and 50-state producer licensing quietly caps the market far below the TAM implied. Every performance figure — 20% savings, 5-hour proposals, “100% 5-star” reviews — is company-sourced in launch week, into a field where Nava (~$90M), Gravie (~$530M), and Gusto/Rippling’s payroll distribution already sit.

A “€40M+” Series B (~$47M) for the Dutch maker of large-area nanoimprint lithography equipment — the “Cypris” platform that mass-replicates the optical waveguides inside AI glasses. The technology is genuinely differentiated (panel-level NIL vs. wafer-level incumbents), but the headline oversells the raise: roughly half is a €20M EIB convertible loan, and most of the equity is state and regional development capital (Invest-NL, BOM, EIC Fund) — only 3M Ventures and Innovation Industries are conventional strategic money, a sign private markets alone haven’t underwritten the scale-up. As an equipment vendor with 10–15 systems placed and no named customers, revenue is its buyers’ capex, and the entire scaling story depends on a consumer AR-glasses market that has already burned $4B+ at Magic Leap — now itself a waveguide competitor — plus Google Glass and North, without ever reaching volume.

A $40M Series A led by Andreessen Horowitz (8VC and Bloomberg Beta returning; HRT Ventures and Next Ladder new) for a startup that runs paid, confidential benchmarks and wants to be the “trust layer” for AI — the Moody’s, S&P, and UL of frontier models. The need is real: buyers lack a trustworthy way to tell which model actually does the work. But the pitch has an issuer-pays hole at its center. The labs Vals grades are the customers who pay Vals, then reprint its scores in their own model cards — and Vals has already disclosed a “customer relationship” with benchmark participants, the exact structure that discredited the rating agencies in 2008. Its confidentiality moat is double-edged: a secret test can be uncontaminated or auditable, rarely both, so trust shifts from “is the test good?” to “do you trust this 25-person startup?” The headline traction — a reported ~$400M valuation and “8× revenue” — is company-sourced and unverified, and the moat is squeezed between free public leaderboards like LMArena (~$1.7B) and enterprises’ own private evals. Selling trust with a model that quietly erodes it.

A $22.25M seed (Khosla Ventures, Cerberus Capital, Caltech; adviser Ion Stoica) for Caltech professor Babak Hassibi’s “Bonsai” family — ternary, 1.58-bit LLMs that shrink Bonsai 2 27B to 5.9GB and run privately on a laptop or phone. The compression is real and the pedigree is genuine, but the story has three cracks. PrismML’s site markets Bonsai as built “from the ground up rather than being compressed from a higher-precision model,” while TechCrunch and Startup Fortune both describe it as compressing Alibaba’s Qwen3 — a contradiction that decides how much durable IP the “mathematical breakthrough” really holds, and whether a Qwen derivative can be cleanly relicensed as Apache 2.0. The headline “98% benchmark retention” and “11M downloads” are company-run and unreproduced. And the structural risk no one has answered: with weights given away free and no pricing, API, or revenue disclosed, PrismML is selling a compression technique into a category that Google (Gemini Nano), Apple, and Microsoft (BitNet) are commoditizing to zero — leaving an exit that leans on an Apple acquisition Bloomberg’s Mark Gurman has publicly doubted.

A $40M Series A (led by Spark Capital, with USV and Lowercarbon; ~$53.5M total) for RELIC — modular GPU pods installed “behind the inverter” at solar farms to run AI on otherwise-curtailed power, skipping the multi-year grid-interconnection queue. Founders William Layden (ex-Cube Hydro, SoftBank Energy) and Varun Palivela (ex-NUVIA/Qualcomm chip architect) are a genuinely well-matched pair, and grid queues and curtailment waste are real problems. But the pitch miniaturizes Crusoe’s five-year-old “stranded energy → compute” thesis on the narrowest, most intermittent power source of the field — and Rune’s own website advertises a 99.5% availability SLA, roughly 44 hours of downtime a year, quietly conceding that solar-tied compute cannot serve the premium 24/7 training market its “AI labs waiting for grid connections” narrative invokes. The unaddressed structural risk: GPUs depreciate on a wall-clock schedule whether or not they compute, so powering them only during curtailment windows means paying full capital cost for fractional utilization — and the headline “85% savings” and “$620M” figures are company-derived, method-free, and name no chip vendor, no customer, and no identifiable 200MW site.

A $41M seed (led by Eclipse, with Craft Ventures and angels including Meta CTO Andrew Bosworth) for “the intelligence of record for the physical economy” — a self-improving industrial AI “brain” plus a proprietary sensing pod, aimed at the “$30 trillion physical economy.” CEO Amir Frenkel is an ex-VP of GenAI at Meta and a real pedigree. The evidence complicates the story: this is a stealth exit built on one WSJ scoop and a single customer testimonial (Steuben Foods), with zero disclosed benchmarks, revenue, valuation, headcount, or model provenance — and the “self-improving brain” is the least testable claim a research-stage lab can make. The structural tell it hasn’t addressed: Frenkel is also Eclipse’s own Chief AI Officer, and Eclipse incubated the company, assembled the team, and led the round — a firm funding its own executive, not an arm’s-length seed. The $41M is one to two orders of magnitude below the physical-AI model leaders, and the sensing-and-reliability niche is already owned by Augury (which Eclipse also backs), Samsara, and Palantir.

A $40M seed (led by DCVC) for a self-styled “frontier AI lab” building “machine-native, composable AI” — models that emit typed decisions instead of text. Its first model, Jev, is waitlist-only. Founder Diogo Almeida has a real OpenAI pedigree (fourth author of the 2022 InstructGPT paper), but the marketing outruns the evidence: the “100X faster and cheaper” headline is graded on the company’s own evals, with reference answers derived by averaging the very models it claims to beat, and critics flagged the latency comparison as measuring Jev against LLMs doing full generation rather than the same constrained decision. “Can’t hallucinate” turns out to mean “can’t emit malformed output” — the founder concedes it can still be “confidently wrong.” The category is already commoditized (OpenAI Structured Outputs, Anthropic tool use, and funded OSS peers Outlines and BAML that raised ~$17M combined for the same problem), and the unaddressed structural risk is capital-shaped: $40M cannot fund frontier-scale training, strongly implying Jev is a small specialized model — yet the base model and compute provider go undisclosed.

A $40M Series A (led by Ribbit Capital, with First Harmonic; $55M total) for “Underwriters Laboratories for AI” — a company that certifies AI agents against its own AIUC-1 standard and then sells insurance priced off that certificate. Founders Rune Kvist (early Anthropic) and Rajiv Dattani (ex-METR COO) pitch “skin in the game”: if a certified agent fails, AIUC pays. The evidence contradicts it. AIUC writes the standard, accredits the auditors, runs the ~5,000-test audit, issues the certificate, and sells the policy — but it’s a managing general agent fronting Beazley’s paper, so the carrier and its reinsurers absorb the claims while AIUC collects commission on a book it graded itself. Two named security researchers have already drawn the precedent: this is the issuer-pays structure that inflated credit ratings before 2008. The unaddressed structural risk: you cannot actuarially price novel, correlated, historyless agent tail-risk with a quarterly snapshot — and the mainstream P&C market is busy excluding the exact risk AIUC is racing to underwrite (ISO CG 40 47/48, effective January 2026). Only one of its seven named logos is tied to a bound policy.

A $40M Series B (led by diagnostics investor Catalio Capital) for at-home vaginal microbiome testing — a $129 swab whose real product is the data behind it. CEO Priyanka Jain frames the 100,000-patient dataset as a women’s health “research engine,” and that dataset, not test margins, is what the round underwrites. The product instinct is real and the market is genuinely underserved, but the clinical-utility flank is exposed: credentialed OB/GYNs say “there is literally no data to tell you what to do with most of the results” and warn of overtreatment, while the “AI infrastructure” moat runs on an undisclosed third-party model. The unaddressed structural risk sits in Evvy’s own privacy policy: it is “not a covered entity” and its data “is not subject to HIPAA” — so the most intimate health data imaginable, retained 10+ years and destined for commercialization, sits under consumer-grade legal protection. The company markets closing the gender health gap while its economics depend on monetizing what that gap produces.

A $5.5M pre-seed (led by Mantis VC — The Chainsmokers’ fund — with SV Angel, BoxGroup, Quiet Capital, Collaborative Fund and 7 others) for a “music software label” from ex-Spotify innovation lead Máuhan Zonoozy that “drops apps the way labels drop records.” The product instinct is real and the founder is a credible repeat operator (Bubbl → Cricket Media, 2018) — but six apps are shipped with zero disclosed revenue, pricing, team size, HQ, or valuation, and the flagship bop is still in beta. The unaddressed structural risk: bop lets users search “its collection of tracks,” remix real songs, and export the derivatives to TikTok and Instagram — activity that legally needs master and publishing licenses, of which VBIS has disclosed exactly none. It’s the same fault line that produced the RIAA suits against Suno and Udio, wrapped in “human, not AI” branding. The closest playful analog, Endlesss, liquidated in 2024.

A $4M seed (led by TQ Ventures, with Neo, Sandberg Bernthal, and CDTM) for Munich-based software that plans, operates, and trades commercial & industrial battery storage — founders who reverse-migrated from Silicon Valley to Germany. The timing is genuine and the sub-megawatt niche is real, but the “first software to combine planning + operation + trading” claim overstates a thin moat (Tesla Autobidder, Fluence, Stem, gridX, and Entrix already do this), and the traction splits awkwardly: “800 companies / 6,000 sites” in press vs. ~100 paying in founder interviews. The unaddressed structural risk: Furo’s +40% uplift and 15→5-year payback depend on intraday price spreads that exist only because storage is scarce — and Furo’s whole mission is to make storage abundant. It sells a return its own success erodes, against rivals with 10× the capital.

A $25M Series A (co-led by Sequoia and SMBC’s Fin Atlas Beyond Fund; $30M total incl. seed, >$100M post-money) for a “Workforce Security Graph” mapping every human, AI agent, and non-human identity to the data it can reach. The pitch is sharp and the Unit 8200 / Talpiot team is real — but the framing (“one graph unifying human + NHI + data”) is nearly word-for-word how Oasis, Astrix, and Token describe themselves, and Sequoia is simultaneously funding all three. Every performance number (“seven-figure ARR,” “74% reduction,” “double-digit customers”) originates with Cymphony and has no third-party validation. The unaddressed structural risk: to map everyone’s blast radius the tool becomes the single highest-value blast radius in the enterprise — yet it never discloses which LLM powers its “Maestro” assistant, whether customer data transits a third-party model, or whether it’s used for training. The vendor that hunts rogue models won’t name its own.

A $10M Series B led by Uber at a $168M valuation (a ~2.7× step-up in eight months) for a ~2-year-old Indian fleet operator that owns/leases ~5,100 vehicles and supplies capacity almost entirely to Uber. The load-bearing fact the coverage skips: Uber is Carrum’s primary customer, its only disclosed equity investor, the lead of both rounds, and a mid-teens shareholder — so no independent party set this price. $168M on ~$24.5M FY2026 revenue and ~$736K net profit implies ~7× revenue and a ~228× P/E for a low-margin, 85–90%-debt-funded leasing business. Every financial traces to the CEO; there is no Uber statement and near-zero independent web footprint. A captive supplier, not an independent company — and its closest rival, Everest Fleet, is bigger, earlier, and backed by the same investor.

A $60M Series A (TQ Ventures; Hanwha, G Squared, JAWS) at a reported ~$350M valuation for a pilotless cargo aircraft that has not yet flown at full scale. The pitch markets a complete spec sheet — 4,000 lb payload, 1,650+ nm range, “a fraction of the operating cost” — when the only flying hardware is a quarter-scale model, making cost-per-ton-mile presently unverifiable. The engineering call is genuinely smart (fixed-wing, proven combustion, no eVTOL hype), but the company is silent on the one thing that decides its fate: an FAA certification path that has stalled Reliable Robotics (~$160M+) and saw Xwing’s autonomy unit absorbed by Joby. $60M is a flight-test budget, not a type-certification-plus-airline budget — and the engine and autonomy suppliers, the safety-gating pieces, are never named.

A $20M unpriced Series B extension (Nimble Ventures, with LM Ventures) for AI that auto-generates CNC machining strategies. Two facts the funding coverage skips: the company’s “$128M raised” claim collides with just $72M of independently tracked funding (Tracxn/Crunchbase) — a gap it has never reconciled — and the 2022 pitch to “deliver autonomous manufacturing” has quietly shrunk to a programming-assist plug-in that lives inside Autodesk Fusion, Mastercam, and Siemens NX. Autodesk is simultaneously a strategic investor and the owner of the platform CAM Assist depends on — free to build native AI CAM and disintermediate it. Add a flagship customer (Lockheed Martin) that is also a repeat investor, and an unpriced bridge four years after the last raise, and the moat looks rentable, not owned.

A $70M round ($65M equity + $5M debt, Qualcomm Ventures) at a $365M valuation, pitched as turning a smart ring into “a computer” — on-device AI and a game-controller mode promised “by the end of September.” The load-bearing facts the funding coverage skips: Ultrahuman is the only major ring maker Oura has kept banned from the US rather than licensed (import ban effective Oct 21, 2025), its Ring Pro workaround clears customs only conditionally, and a US ITC adjudicator found the company falsified evidence about a Texas factory — the same US-manufacturing story now central to its survival and ~45% of its revenue. The “$140M ARR” it tells the press doesn’t reconcile with the ~$68M its India entity last filed. Real, profitable hardware wrapped in a moonshot it can’t afford and a legal fight it hasn’t escaped — against a leader (Oura, $11B) with 80% share and ~30x the capital.

A $7.5M seed (Khosla Ventures, AI House) for a text-only family assistant selling privacy as its wedge to “win the AI assistant race.” The founder’s line is “we’re not sharing your data with anyone” — but Ollie’s own privacy policy routes your household’s email and calendar to three external model vendors (OpenAI, Anthropic, Google) and reserves internal “aggregate” use. The honest version, “we don’t sell it and our vendors promise not to train on it,” is materially weaker than the pitch. Underneath: a calendar+email wrapper on rented intelligence, competing against free, bundled Google/Apple/Amazon defaults (Apple’s on-device model is a stronger privacy story), a cloud-browser login that centralizes standing access to thousands of families’ inboxes, and a launch whose reach was paid-amplified, not organic.

A six-month-old, Unit 8200-founded “context firewall for AI agents” that raised $50M across two seed rounds weeks apart — ~$10M from Sequoia, then ~$40M from Greenoaks. The problem is real, but the category is a shakeout, not a greenfield: Prompt Security sold to SentinelOne (~$250M), Protect AI to Palo Alto (~$700M), and Zenity just raised $125M. Every operating metric — 20+ customers, a 27% filter rate, 17,800 malicious add-ons — traces back to AIR itself, and the CEO concedes the approach “was very easy to copy.” A security vendor sitting inline on all agent traffic, with no visible SOC 2.

A $20.8M seed (Khosla, XYZ) for AI agents that go undercover inside criminal Telegram channels and dark-web forums for the FBI and HSI — where the named agency “customers” are, in Axios’s own words, experimental “short stints,” not contracts, and the three headline deployments are company-asserted through a single paywalled story. The reassurance is “human-supervised, just like an FBI analyst”; the product also does “cyber effects” and “actively contests adversaries.” The load-bearing question — who is accountable when an autonomous agent inside a live investigation engages the wrong person — is one the company has not begun to answer.

A $21M Sequoia studio company promising to “predict outages before they happen” — AIOps’ oldest unkept promise, re-issued with agentic branding and no falsifiable accuracy metric. The core claim is structurally unprovable: when it auto-approves a change and nothing breaks, “prevented an outage” and “there was never a risk” look identical. Conceived, incubated, funded, and CEO’d by Sequoia — an internally-priced round, not a market signal — while the same investor backs a $1B unicorn (Resolve.ai) in the exact same lane. The website sells a modest topology tool; the press sells prophecy.

A $6M seed (SignalFire, Las Olas VC) for a “Harvey for police officers” — a mobile AI that answers cops’ questions on law and policy in real time, in the field. The niche is genuinely open, but it’s the highest-liability corner of AI policing: guidance is consumed before an action, not documented after. The safety pitch rests on an unaudited marketing phrase — “exact quoted answers… with no hallucination” — while a retrieval system can still surface the wrong passage, an outdated policy, or a mis-ranked result, all functionally hallucinations to an officer deciding whether to enter without a warrant. The foundation model and cloud vendor are undisclosed for a tool marketed as CJIS-adjacent, and the company’s own About page contradicts the launch article on two of three founders (the “Harvard Law dropout” hook is absent from the CEO’s bio; the “Boston deputy chief” was a Brookline commander). The buried question — who is liable when an officer relies on wrong AI guidance during a use-of-force or search event — is never addressed. Every headline number (225 agencies, 25 states, “elevenfold” growth) traces back to the company.

A $35M “Series A” from Magna to scale battery swapping across India — except Magna already owns 51%+ and is the only investor, so this is a majority owner topping up its own subsidiary, valuation undisclosed, no external VC pricing the round. The market tailwind is real (India’s gig two/three-wheelers are the one place swapping genuinely beats petrol on cost and uptime), but Yuma is the smaller, later, parent-funded #3+ player: ~80–85% of swaps still come from former parent Yulu, a related party; the “made in Chennai” packs sit on mostly-Chinese cells (India imports ~70–75% of its lithium-ion cells); and the “EBITDA-positive by FY27” claim leans on a metric that excludes battery depreciation and the cost of capital on a 30–50% idle-battery overhang — the two costs that actually break swapping businesses. Market leader Battery Smart already runs roughly 4× the batteries. A committed strategic parent buys time; it doesn’t buy a moat.

A $2.5M pre-seed (Lachy Groom, Together Fund) for a 20-year-old’s autonomous aircraft “engineered to be airborne for over a year” by harvesting energy from ocean wind shear. The maneuver is real physics — dynamic soaring is how albatrosses cross oceans, and Alteon flew autonomous 62 mph loops over the Bay of Bengal. But the flagship promise is unproven by orders of magnitude: the company hasn’t even reached “energy-neutral” flight (its own next milestone), and the best-funded persistent-flight programs on Earth — Google Loon, Facebook Aquila — died despite $100M+ budgets. Worse, the physics is at odds with the mission: dynamic soaring must keep moving through the wind-shear layer to stay powered, so it cannot loiter over a fixed patch of water the way a solar platform can — yet Alteon sells “real-time visibility into their waters,” a station-keeping job for a non-station-keeping tool. A talent bet on a teenager, reported as a technology validation. The record any rival has actually flown is ~64 days.

A $21M+ Series A led by Thrive Capital for an algorithm-free, friends-only photo app from two ex-Instagram engineers — announced eight months after it actually closed (December 2025). The marquee number, “7 million downloads,” is a company-supplied vanity metric its own December 2025 coverage contradicts: Retro then disclosed ~1M users, implying roughly 86% of installs are inactive (a ~14% active-to-install ratio). The privacy-first brand runs on a boilerplate policy that names AWS and Google Analytics and reserves data-sharing with “advertising partners” — against a product built on sharing kids’ photos. And it enters the most reliably fatal category in consumer social: BeReal (~$90M raised, distressed €500M sale), Poparazzi (~$17M, shut down 2023), and Lapse (~$30M, stalled) all died on the same retention cliff and monetization ceiling Retro hasn’t shown it can beat. The honest headline isn’t “$21M raise” — it’s a $21M bet on escaping the friend-photo graveyard.

A $9.5M seed (a16z speedrun, Night Capital, Lightspeed’s Jeremy Liew) for a “factory-direct luxury” marketplace whose entire pitch — “same material, same craftsmanship, coming from the same factories” as Prada and Louis Vuitton — arrives with no named atelier, no authentication, and no chain of custody. The dozen “makers” are generic white-label labels (Boliz, Eix Gnod, Fire Lady Fur); Trustpilot 1-star reviews report goods shipped “directly from China,” “paper-like” quality and a chemical smell. That gap isn’t just trust — fashion-law counsel call implied luxury-equivalence claims a false-advertising and trade-dress liability, and Atorie names the houses in its own framing with zero proof of shared sourcing. Meanwhile the thesis is already owned by Quince ($10.1B) and Italic (~$450M) at roughly 50× the capital, and the headline “$55M run rate” is an 11× projection off ~$5M in company-claimed sales. The demand is real; the provenance story survives no verification.

A $1.8M pre-seed — with no investors named in any source — for EverBowl, a sensor-laden smart feeding station pitched as the “Apple Health for pets.” The promise to “spot when your pup is sick” rests on exactly two anecdotal beta detections across 80+ dogs, while the company itself says clinical studies are only now “informing.” Its headline “97% accuracy” measures the scale, not the diagnosis — sensor precision, not diagnostic sensitivity or specificity. Disease-specific marketing (“tick fever,” “dental infection”) plus bundled telehealth, wrapped in a thin “not a veterinary device” disclaimer, sits exactly where FTC and state-veterinary scrutiny live — and the false-positive window during baseline-building drives the very ER visits it claims to prevent. The company also openly plans to sell health inferences to insurers, a conflict never reconciled with its trust messaging. Good measurement hardware wearing a diagnostic costume — and $1.8M is thin against Fi (~$40M) and Whistle (acquired for >$100M).

The same week WRTN became Korea’s first AI-service “unicorn” at a >₩1 trillion (~$722M) valuation, its first regulatory filing showed a 2025 operating loss (₩58.8B) larger than its entire revenue (₩47.1B). It owns no model — the free portal resells GPT, Claude and Gemini, and the money comes from “Crack”/“OOC” AI character chat, where a small cohort of whales pays with no total spending cap and ~10% of users are minors. Its closest rival, Scatter Lab’s Zeta, is bigger (1.4M vs 550K MAU) and already profitable on a fraction of the capital. A ~21×-trailing-sales valuation priced on unaudited forward guidance, sitting on rented margins and a minor-safety regulatory bullseye. Watch the audited margin, not the top line.

The British Business Bank led a £10M Series A for this Imperial College spin-out (formerly RFC Power) to commercialise a manganese flow battery pitched as the long-duration-storage chemistry that “should win” — >75% efficiency, 20-year life, one-tenth the cost of vanadium. Every one of those claims describes a system that does not yet exist: the company has commissioned only a kWh demonstrator and is still planning its first MWh pilot (in India, with no disclosed order). The tell is the structure — LSE-listed parent Ceres spun it out at a ~£12M valuation while keeping 37% and a revenue share, hedging rather than doubling down. And £10M is a rounding error against Form Energy’s $2B+ in the same race, where Redflow just died for want of factory capital. Promising chemistry; a lottery ticket, not a scaling story.

A $6.8M angel seed priced at a ~$114M valuation (16.8× the raise) — with no named lead, no institution, and not one named angel. Stellaria (formerly the 2019 agriculture startup Farmin) now sells “Stella,” an AI operating system for geospatial intelligence: super-resolution, automated target recognition, maritime tracking. But it owns no satellites, has never named the third-party imagery it depends on, and discloses zero customers or contracts. Its true peers — Orbital Insight, Descartes Labs — were acquired cheaply because software-on-borrowed-pixels rarely stands alone, while constellation owners like ICEYE (~$12B) and home-market Space42 (~$5.1B sovereign contract) own the value. A defense-AI narrative and an unmanaged surveillance-ethics liability, priced by no one nameable. Watch for the first institutional round and the first named contract.

A music-and-media consortium — Universal, Sony, Warner, EA and AMD Ventures — put $76M into the maker of Stable Diffusion, framed by CEO Prem Akkaraju as an “affirmation of our vision.” The framing does a lot of work. There is no lead VC, no disclosed valuation, and the company’s tidy “$232M total raised” quietly resets the odometer after ~$400M of obligations were forgiven in the 2024 restructuring that saved it from insolvency. Meanwhile the architects of Stable Diffusion left in 2024 and built Black Forest Labs (FLUX), now valued at $3.25B — roughly 14× Stability’s implied scale — Midjourney prints ~$500M bootstrapped, and two US lawsuits (Getty, plus Andersen heading to trial Sept 8) sit unresolved atop an off-balance-sheet compute bill the company has never explained how to sustain. Survival capital dressed as validation.

Accel led a $26M seed (with Conviction) for an ex-Yandex/Amazon team building an independent 100-billion-document web index as search infrastructure for AI agents — arguably the deepest index-building résumé of any startup team alive. But “we own the index” is capital, not a moat: Google’s is bigger, the founder himself calls running one “painfully expensive,” and the buyers — OpenAI, Anthropic, Google, xAI, the hyperscalers — are the only players rich enough to build their own. Keenable is selling a shovel to people who own the mine. Compounding it: the whole asset is an unlicensed full-web crawl with silent robots.txt/licensing policies in a post-NYT v. OpenAI climate, and a founder-provenance question (Kremlin-era Yandex search leadership) the company has never addressed. Buy the team; interrogate the index.

Peak XV led a $10M Series A extension (with Arkam and Capital 2b) into a Bengaluru enterprise-voice-agent startup with real logos — CRED, Flipkart, Shell — and credible operator-founders out of Groww and Flipkart. The raise leans on “proprietary AI models and a context graph” as the moat; the company’s own words undercut it. TechCrunch confirms Ringg “currently operates as an orchestration layer routing tasks to different models” because owning the stack “remains too costly” — and it already abandoned model ownership once, as the TTS startup DesiVocal. Worse, lead investor Peak XV also backs Vapi at a ~$500M valuation, a direct rival, and Ringg’s flagship use cases (outbound, collections) sit in India’s highest-scrutiny regulatory lane. Buy the operators and the workflow lock-in; discount the model moat.

Greenoaks led a $37M Series A — with DoorDash and Lightspeed — into a Bengaluru startup whose genuinely novel tail-sitter drone weighs less than its cargo, pitching “cost parity with trucking.” But that pitch is a category error: a 2.2-pound payload competes against a vehicle moving 10,000+ kg, and the one-cent economics live entirely at a 10,000-flight-a-day scale that doesn’t exist — against a still-draft BVLOS regime no Indian regulator has approved at density. The real business is the time-critical medical-sample niche (1,000+ Narayana Health flights); the marketing sells a freight paradigm. Pre-revenue by design, ~$50M raised total — a rounding error next to Zipline’s ~$1.6B, which still must scale ~230× to prove the unit economics nobody in the sector has. Buy the niche; discount the paradigm.

Headline Asia, JIC, and Incubate Fund co-led a ¥2.6B ($16M) round into a Hokkaido University spinout selling small-satellite hybrid thrusters — while promising the “world’s first space-reaching hybrid rocket” by 2028. The thruster business is real, credibly backed, and appropriately capitalized. The launch claim is not: no hybrid has ever reached orbit in 60+ years, the 2028 target rests on a partner (ISC) that has only agreed to “consider” hybrids, and it follows a single ground hot-fire. For scale, Gilmour Space has raised ~$245M on the same architecture and still hasn’t reached orbit — roughly 8× Letara’s total funding. Two companies in one press release; investors are buying the thruster, the marketing is selling the moonshot.

Rhapsody Venture Partners led a $22M SAFE (2× oversubscribed) for an Orlando hollow-core-fiber startup promising light that moves “~47% faster” through air instead of glass — with a $40M hyperscaler order attached. The physics is real, but so is the trap: that 47% is a hard ceiling every hollow-core maker inherits, the identical figure Microsoft’s own Lumenisity published before Microsoft bought it in 2022 and began deploying ~15,000km of HCF across Azure. The biggest buyer of inter-data-center fiber has vertically integrated the exact product. Every fiber is drawn and cabled by Prysmian; Relativity owns no fab, and the SAFE sets no valuation. Credentialed founders (ex-Luminar CTO), the hardest possible moat.

Detroit’s Grounded raised a $5M seed to convert electric and gas-powered fleet vans — the same startup that spent two rounds and a “.evs” domain on “electric” before its base platforms (GM’s BrightDrop, Ford’s electric Transit) were cancelled and it pivoted to combustion. The company now says it was “never a vehicle company,” yet it’s standing up a 50,000 sq ft plant and claiming a “nearly 2,000%” capacity jump on $8.5M total raised — against a partner-slash-rival (Harbinger) that has raised ~$200M. Real demand, credible team; thin, pilot-heavy proof on a chassis it doesn’t control.

Varsity VC co-led a ~$5M seed for a Paris deep-tech promising “30x faster” LLM inference on standard GPUs — a headline number Kog’s own website quietly walks down to “3.5x,” built on a batch-1, 2B custom-model theoretical ceiling its blog admits is “not guaranteed achievable.” Zero independent benchmarks, no named customer, and the ~$5M is secondary-sourced — the Aug 2026 press appears to be the same seed as the Oct 2025 French Tech 2030 label, not a fresh raise. The unanswered question is the moat: Kog sells the exact latency margin that Nvidia’s and AMD’s free first-party stacks are motivated to reclaim.

Peak XV Partners led a $45M Series B at a reported ~$550M valuation — a ~10x markup in under a year for a “faster GitHub Actions” startup. But the hardware edge is now the identical pitch of Depot, WarpBuild, and Namespace, the premium AI layer (“codesmith”) is rented from Anthropic, and it lives atop its largest competitor’s platform. Third-party monitoring tracked 628+ outages, including a 7h18m downtime six days before the raise; the company’s own three surfaces disagree on customer count (3k/5k/6k). A credible operator riding a real wave it doesn’t own.

Asymmetric Capital led an $8M seed for a smart, grid-flexible heat-pump water-heater startup with credible founders (ex-Formlabs, Nebia) — but one that chose the most capital- and labor-intensive path possible, owning both its manufacturing and its own plumbing company, then pricing 3x above incumbents (Rheem, A.O. Smith) that already ship demand-response HPWHs. The grid-revenue thesis is entirely prospective, MCJ’s own post hints the “seed” is really a second seed (~$13M+ total), and every efficiency and savings figure is company-sourced and inconsistent across outlets. A real team, the hardest way in.

GEF Capital Partners led a $93M Series C ($63M equity + $30M debt) into India’s largest shared-EV fleet — but at a ~$170M valuation that reads flat-to-down versus a 2025 mark, with strategic backers Bajaj and Magna both declining to follow on. ~95% of revenue rides quick-commerce demand owned by the very platforms building captive EV fleets to replace it. A real operator riding a boom it doesn’t control.

Lightspeed India led a $9M seed for a two-person startup using AI agents to hunt cooler materials for AI chips — but the moat it describes is a portfolio of patents generated by a frontier model it rents from Anthropic, not owns. Its two flagship proof points are company-attested only, and it enters a field where CuspAI ($450M) and Periodic Labs ($350M+) hold 30–70× the capital. A sharp wedge, not yet a defensible business.

Expedition Growth Capital led a $67M Series B for the 24-year-old Athens conversational-AI company — only its second institutional raise ever, at an undisclosed valuation with no co-investors named. Its “10x ARR to $60M” framing flatters a modest 24-year revenue base, and its bet on a proprietary, non-frontier-LLM stack faces AI-native rivals (Sierra, Decagon, Parloa) valued 40–150x higher. Every headline number is company-reported and unaudited.

Nexus Venture Partners led a $28.5M Series A to build the layer that issues real LLCs, EINs, and payment cards to autonomous AI agents. But a public, code-backed investigation alleges the flagship “Vetta” runtime is an un-attributed fork of Paperclip (MIT, 41K+ stars), the “30,000 customers” figure conflates free signups with revenue, and the defining AML/liability risk of provisioning entities to non-human actors goes unaddressed. A pivot from the founders’ prior AI-SEO startup.

Bessemer and Google’s Gradient Ventures co-led a $10M seed for ex-Spotify engineers bringing real-time recommendation AI to e-commerce. The team is credible and the backers marquee — but the founding credential (“the AI powering 90% of Spotify’s recommendations”) names a system, “Vector AI,” absent from Spotify’s public record, every customer-lift metric is single-merchant and self-reported, and “two-headed Vector AI” is a label, not a moat, in a field of far richer incumbents.

Khosla Ventures and Galvanize co-led a $37M Series B for high-altitude weather balloons and the WeatherMesh AI model, at a reported $250M valuation. The in-situ sensing fleet is a genuinely hard asset — but a WindBorne balloon already cracked a United 737 windshield at 36,000 ft, frontier AI weather models are being given away free, and the company’s market thesis quietly depends on the same public NOAA/ECMWF data it says is dying. Every valuation and revenue figure traces to a single press release.

Saudi Aramco’s venture arm led a $9.5M Series A to scale satellite-and-AI carbon credits from Asia’s rice paddies. The measurement is genuinely strong — Tier-3 dMRV, a Gold Standard methodology, a Nature Conservancy partner — but it sells into a market that voided an estimated 99.9% of rice-methane credits in 2024, and the additionality problem regulators invalidated credits over is behavioral, not something better sensors can solve. Its highest-profile backer is also the archetypal “polluter buys offsets.” Led by Aramco Ventures.

Four ex-Salesforce founders raised a $20M pre-seed — led by Salesforce CEO Marc Benioff’s Time Ventures — for an “AI deployment lab” whose showcase job is fixing Salesforce, while Salesforce’s own Agentforce competes directly. There are zero named customers, no metrics, and “no deck”; the pitch to “use AI to deploy AI” wraps undisclosed third-party models (Claude Code appears in the demo) that its likely suppliers, OpenAI and Anthropic, are now productizing themselves. Closest rival Distyl is already worth ~$1.8B — ~90× June’s raise. Led by Time Ventures.

A $7.9M seed for a crowdsourced design benchmark that pays the frontier labs to generate content, ranks them publicly, then sells the resulting preference data back to those same labs — a triple entanglement it has never addressed. Its headline “~$60M ARR” on a team of ~10 is internally inconsistent with a seed round and independently unverified, its 5.3M “users” are unpaid volunteers with no switching cost, and it inherits wholesale the “funded by those it ranks” conflict now dogging its template, LM Arena ($150M raised). Led by Index Ventures.

Cadre founder Ryan Williams raised $10M+ to sell private-credit managers “one trusted book” — but Ellis appears to run no proprietary model (its own privacy policy hints at an undisclosed third-party LLM), its “$50B+ AUM” describes design partners not platform assets behind only ~5 paying firms, and it quietly re-imports the exact Kushner relationship that shadowed Cadre via Josh Kushner’s Thrive on the cap table. Better-funded twin Maybern has 5× the capital. Led by First Round Capital.

A $13M Series A for voice AI that “breaks the Turing test” — but the company’s own pricing page itemizes ChatGPT 4.0/4.1/5.2 as line items, meaning the intelligence layer is OpenAI and the real IP is a fast STT/TTS shell. At ~$21M lifetime it’s the least-funded named player in a category where ElevenLabs raised $500M at ~$11B, and its 100ms/“Turing” claims are un-audited vendor benchmarks. Led by Seligman Ventures.

The London startup helping MSPs run Microsoft 365 at scale raised $50M from Insight Partners while calling itself “complementary to Microsoft” — but the entire product is a layer on Microsoft’s Graph API, Intune, and delegated-admin model, and the same core job is done free by Microsoft’s own Lighthouse console and the open-source CIPP. Its 300% growth and 800+ customers are company-claimed and unverified, and its $110M funding total doesn’t reconcile with disclosed rounds. Led by Insight Partners.

Martha Stewart’s AI “home intelligence” app runs entirely on OpenAI and Gemini with no proprietary model, promises advice “blind to commercial deals” while affiliate fees are its only disclosed revenue — the exact conflict its own lead investor calls fatal — and launched nationwide asking homeowners to upload insurance policies to hallucination-prone models, with five App Store ratings and no published position on liability. Led by Slow Ventures.

An AI browser from a former Perplexity Comet engineer that acts logged-in as you across the web, claims to “leapfrog OpenAI and Anthropic” on a benchmark it invented and named itself — while renting inference from those same vendors — and ships the field’s most dangerous configuration, autonomous action in authenticated sessions, with no published defense against the prompt-injection attacks that repeatedly broke its founder’s last product. Led by Madrona.

The “World’s Best AI Detector” markets a “1 in 10,000” false-positive rate that the independent University of Chicago audit it cites measured ~19× higher, sees accuracy fall to 73% on the most common real case (human draft, AI polish), and by its own claimed error rate would falsely accuse 5–10% of a student body — all while selling into a market that is banning the category and a CEO who names individuals as cheaters. Led by Menlo Ventures.

A robotics lab that raised the week’s largest seed on a thesis, not a product — letting anyone on the internet drive 100+ physical robots while framing better teleoperation as a path to autonomy it never explains. Founders are elite cyber operators with no robotics background; the technology and enterprise customers are all company-claimed and unverified. Co-led by Index Ventures & Ribbit Capital.

An email-security startup whose own seed investor puts its detection only “on par with Mimecast” — whose alarming market statistics all originate in its own unpublished study, whose three named customers are two of the same three from the seed ten months ago, and which reads every executive email at every customer while naming zero subprocessors. Led by Battery Ventures.

A K-12 “vibe coding” platform whose 700,000-student claim divides by its own teacher count to exactly one class roster, whose binding privacy policy never mentions FERPA and says the service is “not intended to be used by children,” and whose free tier runs on OpenAI credits scoped to a single awareness week. Eight years to a first institutional seed. Co-led by Brighteye Ventures.

A 15% B2B creator-marketplace take rate priced above everyone with the least leverage — while LinkedIn launched a 0%-commission creator marketplace and Substack took sponsorship matchmaking in-house, both six weeks before the round. Its own help docs penalise off-platform payment. Led by Insight Partners.

A private social app sold as having no ads, whose live privacy policy reserves the right to share device identifiers with ad partners — and which told Australian regulators it is not a social media service while telling investors it is. Three engagement metrics disclosed in 2025; none now. Led by Northzone.

An AI that predicts construction projects before the RFP exists — except both prediction features are marked “Soon” on its own site, its integrations are owned by the competitor TechCrunch names, and its privacy policy names zero subprocessors while it sells to firms building airports and nuclear reactors. Led by a16z Speedrun.

Robots that install solar panels, exiting stealth on an elite team and a real DOE prize — but a 48-system fleet target and a 2.8 GW pipeline that imply 7% utilization, a beachhead worth ~1% of project capex, and a data-flywheel moat with no published terms. Led by Obvious Ventures.

Nuclear reactors on barges, headquartered inside a California port — the one state whose fission ban the Supreme Court unanimously upheld, and which DOE’s own national lab flagged as a poor SMR site four months earlier. No NRC filing, no timeline, ~2% of NuScale’s licensing bill. Led by Slauson & Co.

A robot for the six-hour braid. A genuine, prize-winning Harvard prototype meets a black-box tech stack, contradictory TAM math (~$270B vs ~$42B), and an unexamined effect on the deregulated, predominantly Black craft it’s built to speed up. Led by Seven Seven Six (Alexis Ohanian).

“TikTok meets LinkedIn” — an AI agent interviews you on video. The short-form-video wedge is also its largest liability under the EU AI Act, which classifies hiring AI as high-risk and bans the emotion inference a video product invites. 100% on Google Gemini. Led by Luminar Ventures.

Former Webflow CTO Bryant Chou’s AI marketing platform — the website as “your hardest-working employee.” A three-LLM wrapper (Anthropic + OpenAI + Google) with no proprietary model, no public pricing, and its $4B former employer already executing the same thesis. Led by First Round Capital & Y Combinator.

A $66M “seed” at ~$300M to fight Microsoft and Okta — just as those incumbents bundle agent identity for free and the field’s best independents (CyberArk, Veza) get acquired. Pre-revenue, with a Cyberstarts “Sunrise” conflict cloud over its early traction. Led by Cyberstarts.

Malaysia’s profitable WhatsApp inbox. A “169% growth” that doesn’t reconcile with the last public revenue figure, an “AI Agent” rented from OpenAI/Gemini, and total dependence on Meta’s pricing — funding M&A into its two hardest markets. Led by Camber Partners.

AI agents that hire and manage frontline workers over WhatsApp. A "specialised data" moat sitting on an unnamed, rented LLM; algorithmic-hiring bias exposure with no published audit; and Paradox now inside Workday. Led by Dawn Capital.

Europe's record robotics Series A for "generalist" factory robots. Customer-investor circularity, zero disclosed revenue across two record rounds, and NEURA's $1.4B closing one day earlier. Led by CRV with Samsung & LVMH.

The "context graph for enterprise AI." Strong repeat founders, Snowflake as investor-channel-competitor, and a layer the platforms are building, buying, or open-sourcing. Led by Norwest with Snowflake Ventures.

AI agents for in-house legal teams. $40M in six months from Sequoia & Lightspeed, a base-free "40x" growth claim, and a concealed LLM supply chain in software's most capital-flooded vertical. Led by Lightspeed.

Anker-alumni hybrid RV trailer at $120K–160K. All-Chinese/HK capital, China manufacturing in a peak-tariff regime, zero deliveries vs. rivals shipping since 2025, and a trade-secret suit against its CMO.

The $399 "face-to-face game console" from Mirror founder Brynn Putnam. Strong hardware reviews, a category graveyard, no recurring revenue, and a $442.7M writedown precedent. Led by Union Square Ventures.

Airbnb takes 10% of Milan's group-travel leader — and hires away its CEO. ~$580M implied valuation on unaudited revenue, a pay-to-be-recruited coordinator workforce, and a leaderless US launch. Led by Airbnb.

The MIT-licensed AI agent framework that turned down a $20M buyout. 30K GitHub stars, 12.9K forks, zero disclosed revenue, total Anthropic platform dependency, and a documented prompt-injection gap. Led by Valley Capital.

"The Sims meets Twitter" — teens go viral to audiences of AI characters. Real virality, self-reported metrics in the category that produced the IRL fraud case, and teen AI-companion regulatory exposure. Backed by General Catalyst & USV.

Expert network with AI voice onboarding. Eight people, a16z lead, claimed 35K weekly expert signups, zero named customers — and the category's defining MNPI compliance risk unaddressed. Led by a16z.

Marketing strategy, M2C supply chain model, and competitive defensibility at $10.1B valuation. Led by ICONIQ.

AI smart ring for private voice capture. Hardware execution risk, Chinese knockoff threat, and the race against OpenAI/Ive. Co-led by Adjacent & Kindred.

Email infrastructure for AI agents. The SES gap, spam/abuse liability, and prompt injection risk. Led by General Catalyst.

Solid-state transformers for AI data centers. ETH Zurich PhD tech vs. $330M+ in competitor funding. Led by World Fund.

NVMe-native BCDR for MSPs. Datto founder's channel insurgency, active trade-secret litigation, and a trial 6 weeks from funding. Led by General Catalyst.

AI-powered data loss prevention out of stealth. Axonius founder pedigree, Team8 validation, and zero disclosed production deployments. Co-led by Glilot & Team8.

Real-device digital experience testing for emerging markets. IFC/World Bank backing, sovereignty-ready architecture, and revenue-to-valuation gap. Led by IFC & RMB Capitalworks.

Vision AI for urban blight detection. Flock Safety political contagion, unverified privacy claims, and municipal contract risk. Led by Prudence Venture Capital.

"GitHub for hardware test data." Founders Fund preemption, undisclosed Trae Stephens dual role, and defense budget concentration at $1B valuation.

Battery-swap e-motorbikes for East Africa. Spiro's 25:1 scale advantage, FX risk, and grid reliability in Kenya & Uganda. Led by Congruent Ventures.

AI voice agents for M&A due diligence. Accuracy validation gap, MNPI regulatory exposure, and the race against Bridgetown's $19M. Led by Relentless Ventures.

Health data monetization, post-warranty hardware failure patterns, and FDA avoidance at $1.5B valuation. Strategic round.

Data sources, technology stack, and competitive defensibility of the AI-powered recruiting platform at $850M valuation. Led by DST Global.