❯ Modal raises $750 million in Series D as its valuation climbs to $15.75 billion within a year
Valuation more than triples in five monthsModal announced on October 14 that it raised $750 million in Series D at a $15.75 billion post-money valuation. Accel led, Kleiner Perkins joined, and existing investors including Menlo Ventures, General Catalyst, Redpoint and Lux Capital took part. The company says it is on pace to bring in more than $1 billion a year, up more than 15x from a year ago.
Managing GPUs for developersModal is a cloud computing platform. Developers hand it the programs they need to run, and it schedules GPUs for inference, training and code execution, so they don’t have to buy graphics cards or build schedulers; they pay for what they use. To do this, Modal built its own file system, container runtime, scheduler and networking from scratch, drawing compute from more than 20 cloud providers.
Who uses itNamed customers include DoorDash’s AI platform team; Cognition, Ramp and Linear run background agents on it, Runway generates video, Physical Intelligence and 1X train robots, and Suno makes music. On May 21 Modal raised $355 million at a $4.65 billion valuation, when annualized revenue had just passed $300 million; usage has tripled since.
Agents' code needs somewhere to runThe fastest-growing piece is Sandboxes: code written by AI agents runs in an isolated environment so mistakes don’t affect other systems. The product, launched in 2023, now has a $300 million run rate; Modal has started close to 3 billion sandboxes and can run up to a million at once. The more work agents do, the more demand there is for running their code.
Training is nextModal also launched multi-node training: one line of code gets a cluster of high-speed-connected GPUs with no capacity commitment. VM sandboxes are now generally available, and reinforcement learning and post-training products are next. The company has more than 200 employees in New York, San Francisco and Stockholm.
▮ SIGNALEvery extra task an agent takes on means more compute to rent, and selling the “place to run” is becoming as big a business as selling models.
❯ Corporate finance software maker Ramp raises $1.85 billion as its valuation jumps from $44 billion to $60 billion in four months
Second round in four monthsRamp announced on October 15 that it raised $1.85 billion at a $60 billion valuation. Dragoneer led, with Thrive, Founders Fund and others co-leading and Fidelity, T. Rowe Price and others joining; total equity raised is about $5 billion. TechCrunch reports it raised $750 million at a $44 billion valuation in June.
Handling how companies spend and collectRamp sells corporate finance software: company cards, expenses, procurement, supplier payments and collections. It now has 85,000 business customers, up from 70,000 in June, including Visa, Uber, Shopify, Notion and Cursor. Most customers use at least two products, and a quarter use three or more.
Revenue up $500 million in four monthsAnnualized revenue was about $1.5 billion in June and is now above $2 billion, with positive cash flow. Customers’ annualized spending through Ramp exceeds $275 billion, up 175% in a year. The company says its median customer saves 7.5% and does 22% less manual work.
A new business: AI billsRamp launched accounts-receivable automation to invoice, chase payments and reconcile automatically; AI spend management to track spending across model providers in one place; and Router.com, which picks models by performance and price and claims to cut inference costs up to 40%. Early investor Keith Rabois told TechCrunch Ramp will be the main way companies manage AI token spend.
It already holds the booksCompanies’ payment and collection records already sit in Ramp, so adding AI for expenses and collections requires no system switch or data move; another module is far easier than new software. TechCrunch notes that Rippling, which started in HR and payroll, is also moving into cards and expenses, and the two compete for the same finance budgets.
▮ SIGNALWhoever holds a company’s daily bills and payment records has the easiest time selling AI features, and finance software is cashing in on that advantage.
❯ As Anthropic’s revenue growth slows before its IPO, some investors cut valuation expectations to about $1.5 trillion
Expectations come downReuters, citing four people familiar, reported on October 15 that after an investor day at Anthropic’s San Francisco headquarters with CEO Dario Amodei and CFO Krishna Rao, some investors began discussing an IPO valuation of about $1.5 trillion, below earlier expectations of about $2 trillion. The listing is expected in November; the final price isn’t set.
Revenue still growing, just slowerAt the investor day, Anthropic disclosed third-quarter revenue of $16.8 billion, up about 45% from $11.6 billion in the second quarter, and an adjusted profit, which leaves out some costs and isn’t net income. The issue is the comparison: second-quarter revenue had more than doubled from the first, so the third quarter’s slower growth disappointed some investors.
Still deep in the red last yearAnthropic was founded in 2021 by former OpenAI executives including Dario and Daniela Amodei and develops the Claude models. Revenue rose twelvefold in 2025 to nearly $4.6 billion, with an operating loss above $8 billion. Google and Amazon are its investors, compute suppliers and sales channels, and also AI competitors.
A big bill still to payAnthropic has committed to spend at least $518 billion over a decade on cloud, compute and infrastructure. Investors weigh not only growth but whether that spending can be covered, and the growing use of debt to build AI infrastructure has raised their concern about financing costs.
The whole market is coolingThe broader market matters too: AI infrastructure, chip and cloud stocks have fallen, Nvidia-backed CoreWeave’s shares and bonds dropped sharply the same week, SB Energy delayed its IPO in September and Holtec Nuclear pulled its plans. Anthropic’s listing is seen as a bellwether for AI IPOs, with OpenAI’s expected next year.
▮ SIGNALWhen a model company goes public, investors buy not just revenue growth but a huge compute bill signed in advance.
❯ Anthropic offers Blackstone more than 10% interest on debt to rent chips
Double-digit interestAccording to Startup Fortune, citing an October 16 Financial Times report, Anthropic offered Blackstone a coupon above 10% to have Blackstone funds put in $9 billion. These terms are still being negotiated, and no final agreement has been announced.
The money rents TPUsAnthropic’s IPO filing discloses a five-year, $125.2 billion lease for Google’s in-house AI chips, TPUs. To pay for it, Broadcom arranged a $60 billion debt package; a $42 billion senior loan led by banks covers about a third of the lease. Lenders are betting Anthropic keeps paying its rent.
Blackstone takes the riskier layerBlackstone’s $9 billion sits in the junior part, repaid after the senior loan if things go wrong, so it earns more. The report notes similarly rated borrowers pay about 7% in public bond markets, well below what Anthropic is offering.
Borrowing is getting more expensiveRenting chips with borrowed money lets Anthropic expand compute without selling shares, at the cost of a large annual interest bill that weighs more as revenue growth slows. The same week its quarterly growth slowed to 45%, and some investors cut valuation expectations partly over rising financing costs.
Not signed yetHow much Blackstone puts in and the rates and protections for each layer depend on a formal agreement. Broadcom is both chip seller and loan arranger here, and the report says Anthropic’s filing lists that dual role as a potential conflict of interest.
▮ SIGNALThe bottleneck for compute expansion isn’t only whether chips are available, but who will lend and at what interest rate.
❯ Tencent issues $1.8 billion and RMB 26 billion of long-term bonds with terms of up to 30 years
$5.7 billion raisedReuters reported on October 13 (via BusinessDay) that Tencent raised about $5.7 billion in dollar and offshore yuan bonds, above the at least $4.5 billion expected: $1.3 billion of 10-year and $500 million of 20-year dollar notes, plus RMB 20 billion of 10-year and RMB 6 billion of 30-year yuan notes at 2.40% and 3.00%.
Demand four times supplyOrders topped $23 billion, more than four times the final size. Strong demand let Tencent cut pricing: the 10-year dollar note priced 0.9 percentage point over Treasuries, 0.25 point tighter than initial guidance, and both yuan tranches came 0.5 point below initial guidance.
AI spending doublesThe terms say proceeds are for general corporate purposes. Offering documents show first-half 2026 capex of RMB 96.3 billion, more than double RMB 45.9 billion a year earlier, mainly on computing equipment, software and data centers. According to The Standard, it is Tencent’s second bond sale this year after about $4.66 billion in June.
Longer loans, less pressureData centers and compute need years of investment. Money repaid after 10 or 30 years means paying only interest each year, without a big principal bill soon. Moody’s and S&P are expected to rate the bonds A1 and A+, and that strong credit is why the 30-year yuan bond costs just 3%.
▮ SIGNALBig companies are building AI infrastructure like railways: with decades-long cheap debt for an asset that needs continuous investment.
❯ Amazon is in talks to buy AI software company Decart for about $5.5 billion
Negotiating priceIsraeli outlet CTech reported on October 13 that Amazon is negotiating to buy Decart at about $5.5 billion, possibly signing by month-end; Decart did not comment. That is below the up to $7 billion The Wall Street Journal reported earlier.
Making models run on other chipsDecart’s DOS software tunes AI models for a specific chip so they run efficiently on processors from different makers. According to SiliconANGLE, that work used to take months and now takes weeks, letting companies move workloads when chips are scarce or needs change. Decart also builds world models Lucy and Oasis for games and robot training.
A company of under 100 peopleDecart was founded by brothers Dean and Orian Leitersdorf and Moshe Shalev, has fewer than 100 employees and makes tens of millions of dollars, mainly from project work. Its AI-generated game Oasis passed a million users in three days. It raised $300 million at $4 billion in May; Sequoia, Benchmark and Nvidia are shareholders.
Several suitorsCTech says Nvidia made the first serious offer, and Anthropic walked away a month ago after due diligence; SiliconANGLE reported that price at about $6 billion. Decart has worked with Amazon almost since its founding, was among the first to show models running more efficiently on Amazon chips than on Nvidia GPUs, and got early access to Trainium3 last year.
Software for its own chipsAmazon is building its own Trainium chips to rely less on Nvidia, but chips only get used if models run smoothly on them. Buying Decart brings that fitting work in-house. Amazon bought Israeli chipmaker Annapurna Labs in 2015 for about $350 million; this price is more than ten times higher.
▮ SIGNALThe in-house chip race comes down to software: whoever makes models run smoothly on its chips can pull customers away from Nvidia.
❯ SoftBank is in talks to buy security firm Allied Universal for about $18 billion, eyeing robots to replace some guards
Still earlyReuters reported on October 16 that SoftBank is considering buying US security firm Allied Universal at about $18 billion; Bloomberg Law specified that is equity value. Talks are early, the price isn’t set, and SoftBank declined to comment.
America's largest guard companyAllied Universal is the largest US security services company, backed by private equity firm Warburg Pincus and Canadian pension fund La Caisse, with about $12 billion of publicly traded debt. It places guards at businesses, venues and major events.
It also guards data centersAllied’s data center work runs from construction to operation: managing contractors, visitors and perimeters during building, issuing credentials and training during commissioning, and staffing monitoring centers and risk assessments in operation. It says it has served more than 30 million square feet of data center projects since 2014 and uses AI video analytics and GPS patrol tracking.
Robots on patrolSoftBank eventually hopes to replace some guards with robots, in line with Masayoshi Son’s bet on AI and robotics; SoftBank agreed last year to buy ABB’s robotics business for $5.4 billion. A source said humanoid guards are further off.
Buying the workplace firstWhat robot companies lack most is real workplaces and customers. Buying an existing guard company brings a large number of posts and a customer list, including booming data centers, where robots can be tried directly. But patrols, door duty and emergencies must be handled well on site, and whether robots can be more reliable and cheaper than people is still unanswered.
▮ SIGNALCapital betting on robots is starting to buy the industries that employ the most people, so the buyer now sets the pace of replacing labor.
❯ Shareholders approve Fox’s roughly $22 billion purchase of Roku, expected to close in the first half of next year
Shareholders vote yesTheWrap reported on October 14 that preliminary results from the special meetings show the deal approved: about 99.9% of Fox’s voting Class B shares voted yes, and Roku’s final tally will be filed separately with the SEC.
How it's paidThe deal, announced in June, is worth about $22 billion in enterprise value. Each Roku share gets $96 in cash plus 0.9693 Fox Class A shares, about $160 at the announcement price. Roku holders will own about 27% of Fox after closing.
A content company buys the TV entry pointFox has live sports and news plus free streaming service Tubi with 110 million monthly users; Roku runs a connected-TV platform reaching more than 100 million US streaming households, the first screen viewers see. Fox CEO Lachlan Murdoch said the deal pairs Fox’s live sports and news with Roku’s streaming distribution and technology.
Third-largest in the USIn July Nielsen data, Fox had 7.8% of US TV viewing and The Roku Channel 2.9%. TheWrap says the combined company would be the third-largest US media distributor, behind YouTube and the combined Paramount-Warner Bros. Discovery. Fox could sell ads on an entry point it controls, and advertisers could buy content and audience in one place.
The DOJ still has to clear itThe Justice Department issued a second request for information in September; under antitrust law, the waiting period ends 30 days after both companies substantially comply. Closing is expected in the first half of 2027.
▮ SIGNALStreaming competition ultimately comes down to who controls both the shows and the screen viewers see when they turn on the TV.
❯ Workday buys C1.ai to manage which systems employees and AI agents can access
Final agreement signedWorkday announced on October 13 that it signed an agreement to buy C1.ai, without disclosing the price, and expects to close in Q4 of fiscal 2027 (ending January 31, 2027). BofA Securities advised Workday and Qatalyst advised C1.ai.
Who gets into which systemC1.ai, formerly ConductorOne, manages system access for employees and AI agents. It connects email, customer management, code repositories, finance systems and cloud services, granting access when needed, revoking it afterward and logging each step. It also offers tools to find unsanctioned AI use; customers include Ramp, Brex, Instacart and Zscaler.
A six-year-old betCEO Alex Bovee says he and co-founder Paul Querna bet six years ago that identity would become the key control point for companies, and AI proved them right. C1.ai’s research finds 95% of large US companies already run autonomous agents, while Deloitte says only 21% have a mature way to govern them.
HR changes update access automaticallyWorkday is an HR system used by more than 11,500 organizations, including over 65% of the Fortune 500, recording who joins, moves or leaves. After the deal, these changes can trigger access updates directly: old access revoked and new access granted on a role change, without IT working ticket by ticket.
Agents need onboarding tooWorkday already offers an Agent System of Record where companies register and view their AI agents. With C1.ai it will also control which systems those agents can use, managing agents as a kind of employee.
▮ SIGNALOnce agents join a company they have to be managed like employees, and HR software is extending from managing people to managing machines.
❯ Intuit buys billing software company Sequence, reportedly for about $200 million
Deal signedIntuit announced on October 15 that it signed an agreement to buy Sequence, without disclosing the price. MGI Research, citing The Times, put it at about $200 million, a figure Intuit hasn’t confirmed.
Turning contracts into billsSequence turns complex contracts into accurate bills, whether priced by usage, seats or milestones, handling pricing, invoicing, collections and revenue recognition, with AI agents doing daily tasks. The more complex the contract, the more spreadsheets go wrong, which is why customers pay. Its customers are mainly fast-growing software and services companies, including some AI firms.
A London startupSequence was founded in 2021, is based in London and New York, and is led by co-founder and CEO Riya Grover. It raised $19 million led by a16z in 2022. Grover says revenue management has long been the most fragmented workflow in the back office.
Intuit wants mid-market customersIntuit says the deal strengthens Intuit Enterprise Suite and QuickBooks Online Advanced for mid-market businesses, letting it serve industries with complex billing such as B2B software and professional services. Grover will join Intuit’s mid-market product group; Intuit will keep supporting Sequence customers and build integrated products in the coming months.
▮ SIGNALAs usage-based pricing spreads, the small job of billing correctly is becoming a capability finance software companies race to buy.
❯ Berkshire invests another $330 million in Bending Spoons, raising its stake to 4%
$33 a shareBending Spoons filed with the SEC on October 14 that it signed a binding private placement with Berkshire for 10 million new shares at $33 each, raising $330 million, expected to close in the fourth quarter. Berkshire’s stake will rise to about 4%.
Already invested at the July IPOAccording to Latham & Watkins, Berkshire invested $600 million at Bending Spoons’ July 2026 IPO; this is a follow-on. Bending Spoons is headquartered in Milan and trades on Nasdaq as BSP.
Buying apps to fix themBending Spoons buys existing digital products and makes them more profitable by reorganizing teams, rebuilding technology, redesigning interfaces, speeding development and changing marketing and pricing, then buys the next. Its brands include Evernote, Vimeo, WeTransfer, Eventbrite, AOL and Airtable, and it says it has never sold a material business. The filing says AI is often both the goal and the tool of an overhaul.
Money for more dealsProceeds go to general purposes and new acquisitions. It has been busy: it completed the Airtable acquisition on September 4 and agreed on September 10 to buy online whiteboard company Miro for $1.355 billion. For a company that grows by buying, cash on hand speeds the next deal.
▮ SIGNALBerkshire is betting not on any single app, but on a method: buy it, fix it, buy again.
❯ Nscale reportedly plans an IPO at a $35 billion valuation as its Norway data center falls behind schedule
Construction running lateTechstrong IT reported on October 15 that Nscale is preparing to go public at a target valuation of about $35 billion, while its flagship Glomfjord, Norway, data center is about six months late, with 6.4 of a planned 30 megawatts running. Reported causes include cooling leaks, permitting hurdles, construction delays and an on-site fire; a 220-room floating hotel was brought in to house workers.
Selling computeNscale supplies AI training and inference compute to hyperscalers, frontier labs and businesses, building data centers, installing GPUs, connecting power and building its own power plants. It operates sites in Glomfjord and Narvik in Norway, Loughton in the UK and Texas, plus partner sites in Portugal, Iceland and elsewhere. Customers buy usable compute, so revenue starts only once sites are built.
Borrowing before the IPOIn September Nscale announced $3.36 billion of convertible notes led by Third Point, with Nvidia, Apollo funds, Citadel, the Abu Dhabi Investment Council and others, converting into shares at IPO. $2.36 billion arrived at closing; Nvidia’s $1 billion is due mid-November. It says it holds more than $103 billion in contracts.
The UK is short on power tooNscale’s Loughton AI supercomputer, targeted for 2027, needs 90 megawatts, about the use of 315,000 homes. According to UK regulator Ofgem’s June data, 315 data centers wait for grid connections with 73 gigawatts of demand, 60% more than the country’s 45-gigawatt capacity.
Contracts depend on data centersTurning $103 billion of contracts into revenue requires sites built and powered on time. With Norway about a fifth complete and the UK site waiting on the grid, customers start training later and revenue shifts back, which is what IPO investors have to price.
▮ SIGNALHowever large an AI cloud company’s contracts, they only count once each data center is built; construction progress is revenue progress.
❯ Dongshan Precision starts pre-marketing a Hong Kong IPO to raise $2 billion to $3 billion
Hearing passed, pre-marketing beginsOn.cc reported on October 12 that Dongshan Precision passed its Hong Kong listing hearing and started pre-marketing, gauging investor demand before the formal offering. Citing sources, it plans to raise $2 billion to $3 billion.
Regulator approval in handThe Standard, citing Bloomberg on September 29, reported that the China Securities Regulatory Commission approved the H-share plan, with UBS, Haitong International, GF and CITIC Securities as joint sponsors, and the offering could come as early as October.
Circuit boards and optical modulesAlready listed in Shenzhen, Dongshan Precision makes printed circuit boards, precision components, touch panels and LCD modules for global consumer electronics, telecom and auto brands. Citing its prospectus, The Standard says it is the largest PCB supplier for edge AI devices, with 26.9% share in 2025. It entered optical modules last September by buying Source Photonics.
Profit nearly quadruplesFirst-half revenue was RMB 27.8 billion, up about 64%, and profit RMB 2.985 billion, up 2.93 times. The Standard notes its A shares have risen more than 4.5 times since 2025 on AI demand, for a market value above RMB 310 billion.
How it plans to spendProceeds will expand high-speed EML chip and optical module capacity in mainland China, build a high-end PCB plant in Zhuhai for its unit Multek, upgrade existing lines, and fund possible investments or acquisitions, debt repayment and working capital. Optical modules move data by light in data centers; whoever adds capacity first wins more orders.
▮ SIGNALAI is lifting not just chips but the optical modules that connect them, and that capacity has become a reason for manufacturers to raise money.
❯ Salience Labs raises $125 million to expand production of optical switches for data centers
Chip giants investSalience Labs announced on October 14 an oversubscribed $125 million Series B. Bullhound Capital led; AMD Ventures, Intel Capital, SBI, Hudson River Trading and the British Business Bank joined, with existing investors including Applied Ventures and Oxford Science Enterprises.
Switching with lightThousands of GPUs in an AI data center constantly exchange data. The usual approach converts light to electrical signals, switches them electronically, then converts back, wasting power and time. Salience Labs’ silicon photonic switches keep data as light while switching, skipping those conversions, which SiliconANGLE says cuts latency and power significantly.
Out of an Oxford labFounded in 2021 and led by CEO and co-founder Vaysh Kewada, the company builds on more than a decade of research at Oxford and Münster universities. It demonstrated a 64-port all-optical switch at ECOC 2026, calling it a first for silicon photonics; the money funds 128- and 256-port versions and volume production with Keysight and Tower Semiconductor.
How big the market isThe announcement cites Dell’Oro’s forecast that cumulative spending on AI back-end data center switches could approach $1 trillion over 2026–2030; SiliconANGLE cites Cignal AI’s forecast that optical circuit switching will exceed $8 billion by the end of the decade.
More GPUs still get stuck in trafficIf the network can’t keep up, expensive GPUs sit idle waiting for data. If optical switching saves power and time in real data centers, cluster builders will replace some conventional switches, which is why AMD and Intel invest. Going from a show prototype to volume deployment still means clearing manufacturing and reliability hurdles.
▮ SIGNALThe next bottleneck in the compute race is the network, and keeping GPUs from waiting on data is a business worth betting on.
❯ Mueon, founded by former Intel engineers, raises $65 million to develop 3D computing modules
Intel Capital invests tooMueon announced on October 15 a $65 million Series A led by chip fund Maverick Silicon, with Intel Capital, Oregon Venture Fund, IQT and others. The money will hire engineers, refine system design and move toward validation and production readiness.
Stacking chips and memoryAccording to SiliconANGLE, ordinary servers lay processors and memory flat on a motherboard; Mueon’s Cubelets stack them in layers, with wiring, power and cooling designed into the module, which Oregon Venture Fund calls “Legos for the data center.” Shorter data paths mean more speed and less power. The company claims up to 10x better power and cost efficiency, which hasn’t been independently verified.
Two former Intel engineersFounded in 2025, Mueon is led by CEO Wilfred Gomes, who worked on chip stacking at Intel, and CTO Jack Hwang, who helped develop Intel’s transistor etching techniques. It raised a $15.5 million seed round that year. Cubelets can host different chips including GPUs and Arm CPUs, and it is building software so common AI frameworks run on its hardware.
The tighter the stack, the harder to coolCooling is the hard part: processors and memory tolerate different maximum temperatures, a busy processor can overheat memory stacked above it, and heat from upper layers must pass through several layers to escape. Oregon Venture Fund says Mueon solved this but gave no details. Customers will buy only if it beats current options on cooling, reliability and total cost.
▮ SIGNALAI hardware competition is spreading beyond the single chip: how parts are stacked, cooled and powered also decides how much compute gets delivered.
❯ Chip startup Bolt Graphics seeks about $100 million at a valuation of about $700 million
Still negotiatingRuntime Wire, citing an October 15 Reuters report, says Bolt Graphics is seeking about $100 million at roughly $700 million; terms may change and Reuters found no lead investor. Founder and CEO Darwesh Singh confirmed fundraising but didn’t comment on terms.
Not fighting Nvidia on trainingBolt’s Zeus GPU targets scientific computing, chip design and professional rendering rather than AI training. According to IEEE Spectrum, it devotes much of the chip to double-precision compute and is built mainly for path tracing, with roughly half the conventional rasterization performance of a comparable Nvidia card. It comes as a workstation card and a server.
Test chip just powered onRuntime Wire says Bolt announced its test chip powered on two days before the Reuters report, and HPCwire reported hardware-accelerated physics running on it. The product is planned on TSMC’s 5nm process, with delivery estimates ranging from late 2027 to 2028. Customers see a roadmap, not a chip they can buy.
Some skepticismBolt says a 28-GPU Zeus rack matches 280 Nvidia RTX 5090s for real-time path tracing, but Tom’s Hardware called the comparison cherry-picked. Raising at $700 million before production also gives observers pause.
Software is the hard partIf engineers’ familiar simulation and rendering tools don’t run, nobody switches however fast the chip is. Bolt works with Blender, Autodesk and SideFX and will support only specific applications at launch, starting in specialized fields with clear needs.
▮ SIGNALNew chip companies avoid Nvidia’s main battlefield and start in small, specialized fields, where the biggest barrier is software, not hardware.
❯ Labelbox raises more than $100 million as it moves from data labeling to training environments and enterprise agents
Equity plus creditLabelbox announced on October 14 more than $100 million in Series E, including equity and a credit facility, at a $1.25 billion pre-money valuation. Infinitum Partners led, with Wipro Ventures, a16z, B Capital and others.
Profitable this yearLabelbox has more than 400 customers and says it works with over 90% of leading frontier AI labs and hyperscalers. It became profitable this year while growing revenue more than 5x, and runs an expert network of more than 4 million people who write and grade tasks for model companies.
Started with labeling dataFounded in 2018 by Manu Sharma and two co-founders, Labelbox first made labeling tools: people tag images and text to train models. Sharma describes three stages: data, then training environments, now enterprise agents.
Practice grounds for modelsHorizon builds simulated companies full of code, dashboards, spreadsheets and documents where models repeatedly do coding, computer-use and research tasks, scored by programs and experts; the company cites millions of graded attempts. Recursion, launched last week, sells to businesses: agents start from schedules, alerts or tickets, each run is graded, and results improve the next run.
Two businesses feeding each otherLabelbox runs more than 50 agents internally across engineering, research, sales and finance. Models learn work through realistic practice and reliable grading; businesses buying agents care about accuracy. Doing both lets grading experience serve both sides, provided its scores truly reflect whether work was done well.
▮ SIGNALAs AI moves from answering to working, the training data business is upgrading too: it now sells practice grounds and scoring rules, not just labeled data.
❯ Medical AI company Suki raises $90 million to write notes and codes for doctors automatically
Medical AI gets a Series ESuki announced on October 13 a $90 million Series E with new investors Healthier Capital, Crosslink Capital and Hercules Capital and existing ones including Venrock, Flare Capital, Philips Ventures and Zoom Ventures; no valuation was disclosed. The money goes to AI technology, integrations and sales.
A scribe in the exam roomSuki turns doctor-patient conversations into notes, billing codes and other paperwork automatically, so doctors don’t type while listening. It also brings in clinical knowledge from Wolters Kluwer’s UpToDate so doctors can ask questions during visits, and works with any electronic health record.
Revenue doublingThe company says revenue doubled in 2025 and is on track to double again; supported visits are up more than 190% this year. KLAS-validated results show more than $2,600 in average monthly financial impact per provider, with less documentation and after-hours work. More than 450 healthcare organizations use it.
Spreading through partnersSuki sells directly to hospitals and lets other medical software embed it through developer kits; it has more than 30 technology partners, and that revenue grew more than tenfold in two years. When Zoom Ventures invested in 2025, Suki was already integrating notes into Zoom’s healthcare products.
Notes are just the entry pointFounder and CEO Punit Soni says documentation was never the destination, only the entry point; Suki wants to extend into coding, billing and clinical reasoning. Handling these in one flow saves hospitals a tool and a round of data entry, but note accuracy decides whether doctors use it every day.
▮ SIGNALThe fastest way for medical AI to spread is to get inside software doctors already use, not to ask hospitals to buy another system.
❯ North.cloud raises $60 million to help companies see where their cloud and AI money goes
PeakSpan leadsNorth.cloud announced on October 15 a $60 million Series B led by PeakSpan Capital, with Series A lead Companyon Ventures returning and Litquidity, Alpaca, Gaingels and others joining. The money goes to product automation, more integrations and marketing.
All the bills in one placeNorth.cloud gathers bills from AWS, Azure, Google Cloud and Snowflake, finds waste, forecasts budgets and adjusts purchasing automatically. According to SiliconANGLE, its TokenFlow feature also connects to OpenAI and Anthropic models, assigning usage to departments and projects alongside traditional cloud bills.
Started in a garageCo-founder and CEO Matt Biringer spent 13 years selling data center technology before starting North in his garage in 2023. Its first product automated discounted long-term cloud purchases; after seeing companies juggle spreadsheets and five to seven tools, he built a full cost platform. An AI assistant, Noros, explains spending changes in plain language.
$400 million savedIt manages nearly $2 billion of cloud spending, which is customers’ spending, not its revenue, and says it has saved customers over $400 million; customers include AssemblyAI, Tines, Atlan and SnapLogic. It has hundreds of customers and aims for 2,000.
AI bills are getting harder to readAI models charge by usage, and it’s hard to say at month-end which product or team spent what. Putting model costs and cloud bills together by business tells finance what to cut and what to add. Biringer says saving money is a business problem, not just an engineering one.
▮ SIGNALThe more AI companies use, the more they need to know where the money goes, and helping them save on AI has become an AI business itself.
❯ Medical AI certifier Scarlet raises $50 million to expand into the US and Asia-Pacific
Headline leadsScarlet announced on October 12 a $50 million Series B led by Headline, with Notion Capital joining and Atomico, Creandum, Kindred Capital and Entrepreneurs First returning; no valuation was disclosed. Headline partner Nancy Xiao Mclaughlin joins the board.
Medical software must pass it to launchIn Europe, AI medical software needs certification from a designated body before launch. Scarlet is one, saying it is the only one of about 52 that specializes in software and AI; it is also designated by the UK’s MHRA and recognized by the US FDA as a third-party reviewer. Certified reviewers decide; software only speeds the process.
Three years to qualifyFounders James Dewar and Jamie Cox built medical AI at Babylon Health and started Scarlet in 2021, spending three years obtaining designation before issuing a certificate. It charges a subscription covering surveillance, annual audits and reassessment of updates; customers include Flo Health, NHS England, Skin Analytics and AI scribes Heidi and Tandem.
A year savedScarlet says customers have cut a year or more from certification of new AI devices, and update reviews from months to weeks. The money goes to entering the US and Asia-Pacific, adding experts and certifying more forms of medical AI.
Every update means another waitAI medical software updates often and each change may need reassessment. A diagnostic model can be trained in weeks yet wait over a year for review. If Scarlet gets updates through faster, medical AI companies can iterate much faster.
▮ SIGNALThe faster medical AI iterates, the more review becomes the bottleneck, and helping companies through it has become a good business.
❯ Team chat app Zenzap discloses $44 million in total funding and launches AI agents that do work
$44 million in totalZenzap disclosed $44 million in total funding, co-led by Bessemer and Lightspeed with The Garage, to expand development and sales and grow internationally. CTech says the total combines an $11 million 2022 seed and a $33 million Series A, not new money this week.
For non-technical teamsZenzap is a team chat app positioned as a Slack alternative for non-technical organizations. Founded in 2022, it has about 45 employees in Tel Aviv, the US and Europe; CEO Guy Weiss previously founded Stucco Media. It says more than 10,000 teams use it daily across marketing, sales, retail, healthcare, education, construction and food service.
An AI coworker in the chatZenzap launched AI agents that do work: type or speak in a group and they qualify leads, update the CRM, schedule shifts, reconcile payments and track invoices. CTech says an agent takes under 90 seconds to set up, and customers have built more than 40,000 workflows. Admins control who uses each agent, what it can see and which actions need approval.
Easy to start, permissions must be clearUnlike personal chatbots, Zenzap’s agents serve the whole group, where everyone sees what they did, which suits work passed between people. The catch is deciding whose instructions can change systems and whose data the agent sees; Zenzap says it complies with SOC 2, HIPAA and ISO 27001.
▮ SIGNALAgents in team chats are the easiest to use and the hardest to govern, so rules on whose word counts come first.
❯ RapidCanvas raises $29 million to build custom AI apps for companies with a platform plus engineers
Vertex leadsRapidCanvas announced on October 16 a $29 million Series B led by Vertex Ventures SEA & India, with Hitachi Ventures and existing investors Accel and Peak XV, for $52.5 million in total. It is based in Austin, Texas, and led by co-founder and CEO Rahul Pangam.
Selling software and people togetherRapidCanvas builds custom AI apps: engineers learn the customer’s processes, build on its platform and keep running the app, all in one subscription. It targets companies with over $100 million in revenue that have AI budgets but no clear path to results. Pangam says AI’s biggest wins have so far gone to the Fortune 100.
The platform does 60% of the workAccording to SiliconANGLE, the platform handles about 60% of building an AI app in three parts: gathering data and mapping company terms and workflows, ready-made modules for tasks like login and document processing, and checks on reliability, such as whether AI summaries match their source.
Revenue up 8xMore than 75 enterprises in manufacturing, finance and retail have made 1.4 million AI-led decisions with its apps this year; over 40% run several use cases. The company says recurring revenue grew 8x over six quarters. It raised $16 million in 2024.
The hard part is scalingFor companies without AI teams, a provider that does it all is easiest. The difficulty is RapidCanvas’s: if each customer needs more engineers, costs rise with revenue. Hitachi Ventures says it backs the company’s ability to make custom delivery increasingly repeatable.
▮ SIGNALWhen companies buy AI they are also buying people, and only providers that keep shrinking the “people” part can earn software-company margins.
❯ Retail data company Scanntech gets a $180 million Series F investment
Two co-leadsAccording to Latham & Watkins on October 13, Scanntech received a $180 million Series F led by L Catterton, with Partners Group joining as co-lead for its clients and Bradesco and SPX taking part. Scanntech is a Latin American retail data analytics and marketing services platform.
Born at the checkoutAccording to its website, Scanntech was founded in 1992 in retail checkout technology, launched the iPOS system in 2009, entered Argentina in 2010 and Brazil in 2013, is now headquartered in São Paulo, has offices in six countries and more than 500 employees.
700 million receipts a monthIts data comes straight from store checkouts: it says it reads 700 million receipts a month, more than 13 billion in total worth over R$1 trillion, and covered 79% of Brazil’s food retail channel in 2022. That closeness and coverage is what new investors value.
Data has to become a reorder listScanntech sells retailers, brands and distributors tools for pricing, promotions and sales analysis; its retail page lists inventory turnover, out-of-stocks and slow sellers. Retailers buy data to avoid running out or overstocking and to spend promotions well, and it only pays off when stores change orders and prices.
▮ SIGNALThe most valuable retail data isn’t charts but being close enough to the checkout to tell a store what to stock tomorrow.
❯ Reach Security raises more than $50 million to close configuration gaps in companies’ security tools
Series B plus a new productReach Security announced on October 13 more than $50 million in Series B led by Top Tier Capital Partners, with Ballistic Ventures, In-Q-Tel, Microsoft’s M12 and others; no valuation was disclosed. The money funds products, integrations, international expansion and US enterprise and federal growth.
Owning tools isn't configuring themMany companies buy security products but misconfigure them, leaving gaps. Reach reads the live settings of existing tools, determines whether an attack path is blocked or still open, and fixes settings that drift. Setup needs only a read-only API key and no new hardware.
Security veteransCEO Garrett Hamilton and CTO Colt Blackmore built detection and prevention products at Palo Alto Networks, Proofpoint and Cylance. The company earlier raised $20 million led by Ballistic Ventures, with former Palo Alto Networks CEO Mark McLaughlin as an angel. Its customer base doubled in a year and includes Autodesk, Insurity and Ensemble Health Partners.
Inside other toolsThe new Reach Everywhere delivers its findings through an API and an MCP server into tools and AI systems security teams already use, for example reordering a vulnerability backlog by what is actually reachable; fixes go through ServiceNow or Jira and are validated in a controlled environment first.
Automatic fixes must be reversibleAutomatic checks and fixes save repetitive work; analyst firm TAG Infosphere estimated more than 160 hours saved a week. But a wrong rule change can block normal business, so validating first and being able to roll back is what lets teams trust it.
▮ SIGNALSecurity budgets don’t have to go to new equipment; configuring what’s already bought correctly is a business worth paying for.
❯ Process automation company Duvo raises $32 million to enter US retail
$48 million in totalDuvo announced on October 14 a $32 million Series A led by Balderton, with Index, Northzone and Credo Ventures, for $48 million in total; no valuation was disclosed. It hired Michael Green from Databricks as chief commercial officer and former Kmart president Spencer Insolia to run the US, with a former ALDI SÜD CEO as adviser.
Running errands between systemsDuvo learns how work really happens through AI-led staff interviews, screen recordings and documents, then has agents complete approved steps across ERP systems, supplier portals, spreadsheets and inboxes via APIs or existing interfaces. Exceptions and decisions go to the owner, and every approval and change is logged.
Built by a grocery e-commerce teamDuvo was founded in 2025 by the team behind Czech grocery delivery company Rohlik; co-founder and CEO Tomáš Čupr also runs Rohlik Group and knows where retail wastes labor. It has almost 60 employees and customers including UK grocer Morrisons, Holland & Barrett, Boden, Notino and Allegro.
Reconciliation from 70 days to 15At Notino, quarterly reconciliation of about $45 million a year in supplier bonuses fell from 70 days and about 120 specialist hours to 15 days and under 10 hours. At Rohlik, planners used to check about 8 at-risk products per market daily; now agents check every shortage signal around the clock.
Automation without replacing old systemsRetailers run several old systems and staff copy data between them daily. Duvo doesn’t require replacing them; customers can start with one process such as freight or invoice audit and go live in weeks. Čupr says AI should be measured by output per person, not jobs cut.
▮ SIGNALMuch of a company’s repetitive work hides in copying and pasting between old systems, exactly the work agents can take over first.
❯ Mistlabs raises $26 million to move from an AI assistant on Apple devices to its own wearable
Pre-Series ASan Francisco-based Mistlabs disclosed on October 12 a $26 million pre-Series A led by Luminous Ventures and Huaqin Technology to develop a standalone AI wearable; CEO is Peter Mo.
An assistant that thinks aheadIts product, Ambi, left beta on iPhone, Apple Watch and Mac. According to SiliconANGLE, it learns about users from conversations its devices hear, understands their work style better over time, and lets a task move between Mac, phone and watch. Audio storage is off by default and users set what it can do.
Working during the meetingIts Routine feature hands a recurring job to several agents. In the company’s example, during a competitor meeting Ambi fact-checks in real time, compares rival products and drafts a proposal on the MacBook before the meeting ends, then keeps tracking the project and flags missing information and new developments.
Next: hardware and AI negotiationIt plans its own wearable later this year, with a waitlist open, to give Ambi more context about users and surroundings. In coming months, two users’ Ambis in the same meeting will be able to share context within permissions, find shared priorities, flag disagreements and propose outcomes both accept.
Why wear one more deviceThe more an assistant knows, the more it helps, but that also means it hears more, so users must decide what it may know. A wearable has to offer help phones and watches can’t, or people have no reason to wear one more thing.
▮ SIGNALThe challenge for AI wearables was never whether they can be built, but whether they give people a reason to wear one more device.
❯ DeepSeek plans to lead a round in Winverse, a robotics AI company founded by former Tencent scientist Hu Han
Seed round not finalizedIfeng Tech, citing Bloomberg, reports that DeepSeek plans to lead Winverse AI’s seed round of tens of millions of dollars, with Meituan’s Longzhu Capital also committing, valuing it at about $400 million afterward. The deal isn’t final; Hu Han, DeepSeek and Longzhu did not comment.
Building brains for robotsFounded in 2026 and still in stealth, Winverse AI builds AI for robots, known as embodied intelligence: helping AI understand its surroundings and control machines to complete tasks.
A founder from TencentFounder Hu Han was Tencent’s chief scientist leading vision-language and embodied foundation model teams; he holds a PhD from Tsinghua University and previously worked at Microsoft Research Asia and Baidu.
DeepSeek bets on robotsThis is among DeepSeek’s largest startup investments. It earlier took a 2% stake in Unitree through its Shanghai listing to work on embodied models. If the deal closes, Winverse may get not only money but a chance to work with DeepSeek’s model team.
▮ SIGNALModel companies are starting to back robotics teams, and model capability is looking for a body that can do physical work.
❯ Binance.US is in talks to raise up to $100 million at about $1 billion, far below its 2022 valuation
Still in talksCointelegraph, citing Bloomberg, reported on October 15 that Binance.US is seeking up to $100 million at about $1 billion to expand and win back market share from larger rivals. Amount and valuation aren’t settled.
$4.5 billion four years agoBinance.US announced a seed round of more than $200 million in 2022; according to Cointelegraph, its pre-money valuation was $4.5 billion. A $1 billion deal now would be less than a quarter of that.
A US crypto exchangeBinance.US offers crypto trading to US users. According to its help page, it is operated by BAM Trading Services using licensed matching and wallet technology.
Regulatory historyCointelegraph notes that in 2023 Binance Holdings settled with US authorities for $4.3 billion in penalties plus added oversight; then-CEO Changpeng Zhao pleaded guilty, served four months and was pardoned by Trump in 2025. Binance.US told World Liberty Financial, co-founded by the Trump family, about the round, but it won’t participate.
Repriced as it is todayA valuation far below four years ago shows new investors are pricing the exchange’s current business, not past expectations. If the round closes, Binance.US gets money to expand but will have to win back share and valuation through trading.
▮ SIGNALA drop from $4.5 billion to $1 billion shows crypto investors no longer pay for past stories, only for the business as it stands.
❯ Chariot raises $10 million to use AI to find consumer cases and organize evidence for law firms
A law-firm service provider gets a Series ACrunchbase News reported on October 15 that Chariot raised $10 million in Series A led by Silverton Partners, with Lightbank, FJ Labs and others, for $15.7 million in total; no valuation was disclosed. It is breaking even, and the money mainly goes to hiring and product, including AI tools for case review.
Finding plaintiffs for law firmsWhen companies add hidden fees, make subscriptions hard to cancel, track users without consent or leak data, many consumers are hurt but each loses little, so few sue. Chariot scans public records such as federal filings for potential cases, uses AI to analyze allegations and write memos for firms, finds affected consumers through ads and fintech partners, checks eligibility in real time and sends vetted claims to partner firms.
200,000 submissions a monthBased in New York with about 15 people, Chariot typically gets more than 200,000 claim submissions a month, tests ads for about 30 potential cases monthly and pursues about 10, and has taken on about 80 cases since adopting its model. It charges firms flat fees while firms work on contingency; it isn’t a law firm and doesn’t sue itself.
A new nameChariot was formerly ClaimClam, which announced a $3.6 million seed round led by Indicator Ventures in April 2025. Founder and CEO Zim Hang says the goal is to “make unethical unprofitable.” It has received cease-and-desist letters over using company names in ads.
Complete files make cases easierIn mass consumer cases, checking thousands of people’s eligibility and evidence takes the most time. Chariot plans to let users connect email so software finds receipts and charges automatically; Hang estimates that could raise claim volume 10x to 100x, letting firms take more cases and decide faster.
▮ SIGNALAI makes small, scattered consumer claims worth pursuing, and doing the legwork of organizing evidence for law firms has become a paid business.
OUTLOOK
More money, delivery has to keep upThis week’s funding and acquisitions all come down to delivery. Compute companies have to turn money into finished data centers, connected networks and GPUs delivered on time, and Nscale’s delay in Norway is a reminder. Application companies have to prove AI really does work for finance teams, doctors, security teams and law firms, and that customers keep paying.
Buying is faster than buildingWorkday, Intuit and Amazon all chose to buy existing companies to add access management, billing and chip software rather than build from scratch. Next, watch whether these companies deliver on time, whether customers renew, and whether borrowed money can be covered by revenue.