The Cost We Choose

Kanurra

Most employers are asked to judge a pharmacy benefit from the invoice. Kanurra wants them to inspect the machinery that produced it.

A prescription claim can hide several businesses inside one number: the pharmacy payment, the plan bill, the manufacturer rebate, and the fees collected along the way. The employer carries the cost, but the economic chain is often treated like a trade secret.

Semih Gultekin built Kanurra around a different agreement with the buyer. The New York pharmacy benefit manager says it charges one fixed fee per employee, passes through the pharmacy cost without spread, and credits 100% of rebates it receives back to the health plan. Every claim is supposed to open into its parts.

Kanurra has announced $6.5M in financing from Necessary Ventures, Asylum Ventures, Daybreak Ventures, Virtue VC, Ford Street Ventures, and Browder Capital. An earlier founder disclosure described $6.35M from the same core group, so the current company release supports the rounded headline while the earlier figure stays in the record.

A 2025 FTC staff report found that the Big 3 PBMs' affiliated pharmacies marked up many analyzed specialty generics by hundreds or thousands of percent. Kanurra is entering that argument from the smaller-employer side, where pharmacy benefits are often bundled into a larger plan and the people paying the bill have the least leverage to unwind it.

Gultekin says the capital will expand covered lives and build infrastructure for partners and enrollees. The hiring plan spans pharmacists, implementation, operations, broker and TPA sales, and engineering. That list tells the truth about the work: a PBM is a regulated service carrying clinical decisions, data, network relationships, member support, and an uneventful go-live.

Kanurra says its AI can approve clean prior-authorization cases while a licensed clinician owns anything off-pathway, and that AI never issues a denial. Its website also publishes modeled savings examples. Neither is independent operating proof yet. Public sources do not name customers, report covered lives, or provide audited savings.

Kanurra's investors are funding a useful challenge to the PBM status quo: show the buyer where the money moved, then accept responsibility for the system that moved it. The next evidence will come from implementations where the claim ledger, clinical handoff, employer economics, and member experience all have to agree.

Deeper Analysis 👇

https://devcuration.com/articles/kanurra-raises-6-5m-in-funding

Aptatek Biosciences, Inc.

PKU care can turn one drop of blood into a logistics problem.

The fingerstick may happen at home. The sample still travels to a centralized laboratory, and the person managing phenylketonuria waits for the number that helps a care team adjust diet or medication.

Aptatek Biosciences is financing a different handoff. The Princeton clinical-stage diagnostics company has announced the initial close of a round led by Canterbury Scientific Ltd. and Longview Innovation. Terms were not disclosed.

The money will support the clinical evaluation and continued development of PheCheck, Aptatek's investigational system for quantitative phenylalanine monitoring at home. It combines DNA aptamer sensing, a single-use cartridge, a reusable handheld reader, and a smartphone-connected application that can share measurements with authorized care teams.

That product architecture matters because the hard part is not moving a laboratory result onto a phone. Aptatek has to bring the measurement closer to the patient while preserving the analytical trust that makes the result useful.

PheCheck has FDA Breakthrough Device designation. The system is now in a registered clinical study, NCT07551921, that the company expects to complete before year-end as it prepares for a planned FDA submission.

The boundaries deserve equal billing. Breakthrough designation is not approval. PheCheck remains investigational, has not been cleared or approved by the FDA, and is not commercially available. The clinical data still have to carry the argument.

Michael Boyce-Jacino is leading a company built from DNA aptamer work developed by co-founders Tilla Worgall and Milan Stojanovic at Columbia University Medical Center. Their science is designed to detect small-molecule biomarkers that are difficult to measure with antibodies outside the central lab.

The investor mix follows the work. Canterbury Scientific brings diagnostics manufacturing and controls experience. Longview backs university-derived hard science and has supported Aptatek as the research moved toward a regulated product. Aptatek previously raised a $4M Seed in 2022, then added undisclosed financing tranches in 2024 and 2025.

This initial close is capital for the stretch where a clever assay has to become a dependable clinical system. Patients do not experience decentralization as a platform strategy. They experience it as less time between providing a sample and understanding what to do next.

That is the promise Aptatek is now paying to prove.

Deeper Analysis 👇 https://devcuration.com/articles/aptatek-biosciences-holds-initial-close-of-financing

Jeeves

Stablecoins are starting to disappear inside the product. That may be the most important part of Jeeves' $110M equity round.

Dileep Thazhmon and the Jeeves team are not asking a finance department to become a crypto operation. They are asking it to manage cards, accounts payable, treasury payments, expenses, and cross-border money movement through one system while stablecoins do more of the settlement work underneath.

CoinFund led the financing, with AllianceBernstein, Andreessen Horowitz, Coinbase Ventures, CRV, GIC, Global PayTech Ventures, ParaFi, Vista, Wintermute, Y Combinator, and others participating. The round was announced September 29, but Jeeves did not disclose a formal series label or a new valuation. Its $2.1B valuation belongs to the company's 2022 Series C, not this transaction.

The operating numbers explain why David Pakman and CoinFund leaned in. Jeeves says annualized platform volume has passed $5B after tripling year over year. Stablecoin-settled volume went from effectively zero to $1.5B annualized in 8 months. Revenue grew 4x in 14 months, and more than 80% of customers use multiple products.

Those are company-reported metrics, but the direction is useful. Enterprises rarely wake up wanting a new settlement rail. They want a supplier paid faster, cash visible across currencies, expenses reconciled, and controls preserved while the company enters another market. The rail earns its place when the workflow gets simpler.

Jeeves is putting the new capital behind that handoff. The company is launching a proprietary stablecoin wallet with payouts to 190 countries, expanding its stablecoin card offering from 25 to 35 countries, adding global AI spend tracking and accounts receivable, and opening an office in Madrid.

There is real complexity under the cleaner interface. Jeeves' legal notices say licensed third-party providers execute money and digital-asset movement while Jeeves administers the technology platform. Every new country adds another mix of banking relationships, payment partners, currencies, compliance rules, and failure points that customers never want to see. The promise lives or dies in those handoffs, far from the pitch deck.

That is the work behind the round. Jeeves has to make a fragmented financial network behave like one operating system without pretending the fragments went away. If it can keep the controls intact while money moves faster, stablecoins may win enterprise adoption by becoming the least interesting thing on the screen.

Deeper Analysis 👇 https://devcuration.com/articles/jeeves-raises-110m-in-equity-funding

Protego Ventures

The longest distance in defense technology is often between a system working under pressure and an allied institution being able to buy, integrate, and renew it.

Protego Ventures has closed Fund I at $125M to finance that crossing for early-growth Israeli defense companies. The vehicle plans to write $5M-$15M checks into autonomous systems, AI, cybersecurity, sensing, drones, resilient navigation, and critical infrastructure technology.

Lital Leshem and Lee Moser built Protego around a specific market failure. Generalist investors can struggle with the operating logic of defense companies, while defense capital can miss the relationships and technical context inside Israel's security ecosystem.

Protego is selling more than capital. Its team combines venture investing, military operations, diplomacy, and access to US and allied defense networks. The pitch is that battlefield evidence can become a commercial advantage only when somebody knows how to carry it into repeatable contracts.

Fund I already has a useful test case. Protego says it led XTEND's $70M Series B, and XTEND has since listed on the NYSE under XTND. The portfolio also includes ASIO, Prisma Photonics, and Rilian across navigation, infrastructure sensing, autonomy, and agentic cyber defense.

The fund economics contain an unusually candid choice. Protego's site had described a $150M vehicle, but the firm stopped at $125M. Lital Leshem told TechCrunch that XTEND could become a fund-level return driver, and raising more money would dilute that potential across a larger capital base. Bigger was available. Concentration looked better.

TechCrunch also reports that executives from Ares Management, rather than Ares Management itself, were among the largest backers and supplied a $30M LP commitment. That distinction matters because capital reporting gets sloppy fast when a recognizable institution sits near the sentence.

Protego is already discussing Fund II and a broader US investment footprint. The work ahead is less glamorous than the headline: help Israeli founders translate operational urgency into procurement credibility, production capacity, compliance, and durable customer relationships across allied markets.

Defense technology is having its venture moment because governments need faster answers. Protego's $125M close puts money behind the harder part, making sure the companies with answers can keep delivering after the emergency becomes a budget, a program, and a contract renewal.

Deeper Analysis 👇

https://devcuration.com/articles/protego-ventures-closes-125-million-fund-for-israeli-defense-tech

Maven Robotics

A store can change its order in 48 hours. The pallet still has to be rebuilt box by box before the truck can leave.

That handoff from digital demand to physical work is where Maven Robotics has chosen to begin. The Santa Clara company has raised a $100M Series A to scale an industrial robotics system built around complete workflows, starting with mixed-case palletizing and tote handling.

RoboStrategy led the round. LocalGlobe, Vine Ventures, XTX Ventures, Shorooq, and Presight also participated. Maven plans to build 250 third-generation robots and begin designing a fourth-generation platform.

The machine is not pretending to be a person for the camera. It moves on a wheeled base, uses 2 arms to handle loads up to 30 kg, and connects with warehouse-management systems so varied cases arriving from different factories can become store-specific outbound pallets.

That job is less glamorous than a humanoid demo and far less forgiving. The box mix changes. Surfaces and weights vary. The robot has to move at production speed, place the right case in the right position, work safely around people, and keep doing it across shifts. A customer buys the outcome, not the vocabulary around physical AI.

Maven is measuring work finished inside the customer's operating day.

Hamza Derbas, co-founder and CEO, spent 9 years in Apple's special projects group after earlier automotive and EV work. Khalid Derbas, co-founder and CFO, brought a private-equity background. They started Maven in 2024 and made a useful strategic choice: go deep on one customer problem, clear the operating bar, then carry the platform and data into the next task.

Maven says as many as 8 robots are already working 16-hour days with 99% or higher uptime, including deployments with an unnamed Fortune 250 consumer-goods company. Those are company-reported figures, but they point to the test that matters. Maven has to turn early production proof into repeatable deployments without sanding away the flexibility that makes the system broader than one machine for one job.

The $100M gives that experiment more surface area. Maven can manufacture more systems, collect more operating data, and push from palletizing and tote handling into broader material handling and assembly. Each new workflow will ask whether the same architecture can learn a new job without rebuilding the company around it.

Maven now has enough capital to build the fleet that tests its thesis in public. The next pallet is still specific, physical, and due before the truck leaves.

Deeper Analysis 👇 https://devcuration.com/articles/maven-robotics-raises-100m-industrial-robots

General Compute

Finn Puklowski is using debt capital to answer a question the chip industry usually leaves to benchmarks: who will actually buy the alternative hardware and put it in front of customers?

General Compute has signed a multi-year agreement with Cerebras to deploy wafer-scale inference at scale. Capacity is expected to open in Q1 2027, with agentic coding as the first workload.

The contract value, system count and capacity commitment were not disclosed. General Compute did disclose that this is its largest hardware commitment and the first deployment drawn against the up to $400M debt facility it secured from Upper90 in July.

AI accelerators do not reach customers on architecture alone. Somebody has to buy the systems, operate the fleet, package the service and sign contracts software companies can use.

Nvidia's advantage travels with a mature cloud, financing and deployment ecosystem. Alternative chips can post extraordinary benchmarks and still wait outside the market because buyers do not want another balance-sheet problem.

General Compute is trying to become the deployment arm for that alternative market. The company will finance and operate Cerebras systems, then sell dedicated inference capacity under one contract and one set of SLAs.

The first customer thesis is sharp. A coding agent may call a model hundreds or thousands of times while it reads a repository, plans a change, edits files, runs tests, studies the failure and tries again. Each call sits on the critical path. Token latency accumulates into the developer's waiting time.

Cerebras attacks that constraint with wafer-scale processors that keep large amounts of memory close to compute. CEO Andrew Feldman and CTO Sean Lie have spent a decade building the category. General Compute co-founders Finn Puklowski and Jason Goodison bring the financing and operating path, giving that engineering another route into production without requiring every coding-agent company to own the machine.

The open questions matter. No deployment-specific economics, utilization commitments or customer reservations were published. Q1 2027 is a schedule, not installed capacity.

Still, the agreement moves the alternative-chip argument beyond a performance chart. General Compute is committing capital, Cerebras is committing systems, and coding-agent builders are being offered a faster tier without inheriting the hardware. The next proof arrives when those first racks turn financing into working tokens.

Deeper Analysis 👇 https://devcuration.com/articles/general-compute-brings-cerebras-speed-to-coding-agents

Varda Space Industries

A laboratory becomes more valuable when it can run the next experiment before the last result turns into institutional memory.

That is the industrial bet behind Varda Space Industries' $251M Series D. Lux Capital and Natural Capital led the round, with Founders Fund, Khosla Ventures, Caffeinated Capital, General Catalyst, 8090 Industries, Giant Step, and Also Capital participating. Varda says the financing brings total capital raised to $598M.

Will Bruey and Delian Asparouhov did not start Varda simply to put a factory in orbit. They built the return trip into the product. The W-Series spacecraft processes materials in microgravity, separates from its satellite bus, survives reentry at roughly Mach 25, and brings the payload back for analysis on Earth.

That last part is where space manufacturing stops being a demo and starts behaving like infrastructure. A pharmaceutical partner needs more than an interesting crystal from one heroic mission. It needs another run, a controlled process, a recovered sample, comparable data, and a schedule that can sit inside an actual drug-development program.

Varda has now completed six successful reentries since its first mission launched in 2023. The company says more than a dozen additional launches and returns are planned through 2028, while an agreement with Southern Launch covers 20 returns to Australia's Koonibba Test Range through the same year.

NASA has spent decades studying how microgravity can produce larger, more ordered, or more uniform crystals that may improve drug research and formulation. The commercial question is narrower and harder: which molecules benefit enough to justify leaving Earth, and can those results survive manufacturing controls, regulation, and patient use?

Varda is funding the loop required to answer that question repeatedly. The company says the new capital will increase mission cadence, deepen pharmaceutical partnerships, and move it closer to delivering medicine manufactured in space.

Reuters also reported a $1.6B valuation. The number is attention-grabbing, but the useful signal is what investors are underwriting: aerospace reliability, pharmaceutical patience, and enough repetition to make orbit feel less like an expedition and more like a processing step.

The next milestone will not be another beautiful image of a capsule under parachute. It will be a pharmaceutical team treating the next Varda flight as part of its normal development calendar, with the previous result already shaping what goes up.

Deeper Analysis 👇

https://devcuration.com/articles/varda-raises-251m-space-drug-supply-chain

Listen Labs

A CRM can tell you that a customer churned, ignored an offer, opened a support case, or stopped using a feature. The expensive question arrives one second later: why?

Salesforce has signed a definitive agreement to acquire Listen Labs, an AI customer-research company built around closing that gap. Terms were not disclosed, and the transaction is expected to close in Salesforce's fiscal Q4 2027 after customary conditions and required regulatory clearance.

Listen Labs does more than automate a survey. Its agents design studies, recruit participants, run interviews, synthesize findings, and simulate customer responses. Salesforce says the platform can draw from more than 50M participants across 120+ languages.

That puts the acquisition upstream from the usual CRM action. Marketing Cloud and Service Cloud already sit where companies campaign, respond, retain, and sell. Listen Labs can help explain the human reasoning those systems are reacting to, then turn it into context for the agents operating inside them.

Alfred Wahlforss, CEO and co-founder, and Florian Juengermann, CTO and co-founder, began the company in a San Francisco hacker house in 2023. A hackathon project became an AI interviewer after their prior app attracted 20,000 users overnight. The commercial platform launched in April 2025.

By January 2026, the company said it had interviewed more than 1M people, served hundreds of enterprises, grown annualized revenue 15x in nine months to eight figures, and raised a $69M Series B led by Ribbit Capital. Evantic joined returning investors Sequoia Capital, Conviction, and Pear VC, bringing company-reported total equity funding to $100M.

The founders were raising another round when they met Marc Benioff. They chose Salesforce distribution instead. Wahlforss remains CEO, Juengermann remains CTO, and the team says it will keep building Listen inside Salesforce AI Labs.

The fit is obvious enough to sound simple, which is usually where integration gets difficult. Customer interviews are evidence, simulations are models, and enterprise agents still need people to decide when each deserves trust. More context can improve a decision, but only if teams preserve the difference between what a customer said, what a model inferred, and what the business chose to do.

Salesforce is buying the part of the customer record that rarely fits inside a field. Listen Labs now has to make that context travel across a much larger enterprise without losing the human signal that made it valuable.

Deeper Analysis 👇 https://devcuration.com/articles/salesforce-to-acquire-listen-labs-for-ai-customer-research

Ray Therapeutics

By the time many people with retinitis pigmentosa reach a vision-restoration trial, the photoreceptors that once caught light are already gone. Ray Therapeutics is asking the surviving retina to take over the job.

The Berkeley biotech received a $7,975,224 grant from the California Institute for Regenerative Medicine to support clinical development of RTx-015. CIRM approved grant CLIN2-17086 on April 24, 2025.

This is public grant funding, not an equity round. It supports Phase 1 ENVISION, where Ray has to learn whether one injection can safely produce a light signal that becomes useful vision.

Paul Bresge, co-founder and CEO, is the principal investigator named by CIRM. Ray uses gene therapy to deliver a light-sensitive protein to retinal ganglion cells. Those cells can survive after rods and cones have degenerated, creating a possible route around the many mutations associated with inherited retinal disease.

The biology is clever, and the clinical job is unforgiving. A retinal cell responding to light is only the first handoff. The signal has to remain safe, strong and stable. The brain has to interpret it. The trial then has to measure whether any change gives a person something useful outside the clinic.

ENVISION is an open-label, non-randomized, dose-escalation study in adults with retinitis pigmentosa or choroideremia. ClinicalTrials.gov now lists it as active but not recruiting, with 10 participants enrolled and five years of follow-up planned.

CIRM had already awarded Ray two translational grants totaling almost $8M. This award brings the agency's support to $15,973,707 across three grants. The earlier money moved the science toward the clinic. The newest award pays for the part where it has to answer to patients, protocol and time.

Ray has added serious momentum since then. RTx-015 received FDA RMAT and EMA PRIME designations in April 2026. Ray also closed a $125M Series B to support later-stage development and commercial readiness.

Those milestones strengthen the development path without completing it. Regulatory access, private capital and a public grant each solve a different problem. None can substitute for the evidence carried by the people in ENVISION.

Vision restoration sounds cinematic until it becomes dose escalation, safety monitoring and years of follow-up. The real work is quieter: moving a signal from a bioengineered protein through the retina, into the brain and back into a person's day where seeing something has to become being able to use it.

Deeper Analysis 👇

https://devcuration.com/articles/ray-therapeutics-8m-cirm-grant-rtx-015

Relay

International Paper just did something more useful than writing Relay a testimonial: it led the company's $36M financing.

That choice carries weight because International Paper is not learning about Relay from a pitch deck. It has used the platform inside loud, demanding industrial environments where a missed handoff can become downtime, rework, a delayed repair, or a safety problem before anyone gets around to updating the system.

Chris Chuang, Relay's CEO and co-founder, has spent years turning a connected radio into an operating layer. Workers keep push-to-talk hardware while Relay adds translation, location, safety alerts, workflows, and a data trail.

The useful detail often arrives out loud because the person who notices it is moving, wearing gloves, responding to noise, or solving the problem before a form can ask the first question.

Relay's company-sponsored survey of 500 manufacturing professionals found that 57% said their most important operational communication happens aloud and in person. Only 17% said almost everything workers notice reaches a system, log, or formal report. Relay calls the missing portion signal evaporation.

The new capital is meant to capture more context without turning frontline work into data entry. Relay plans to expand its AI platform, enterprise integrations, safety tools, hardware, and an agentic layer for querying shift data and acting on it.

New investors include clients of Cerity Partners Ventures, Harmonic Growth Partners, and Thayer Investment Partners. G2 Venture Partners and Wind River Ventures increased their positions. Relay says total capital is now above $90M; it did not disclose a round label or valuation.

Relay says it serves hundreds of thousands of workers across nearly 10,000 sites, reaches almost 10% of the Fortune 500, and processes more than 1B data points weekly. Those are company-reported figures. Axios reports about 300 employees and roughly 100 planned hires.

The interesting bet is bigger than replacing a radio. Relay has to preserve the speed of a spoken warning while giving that warning enough structure to reach maintenance, safety, staffing, and production systems before the shift moves on.

Customer-led capital gives Relay money and a demanding operating environment in the same relationship. The next layer will be judged in the minutes between someone noticing a problem and the organization being able to act on what was said.

Deeper Analysis 👇

https://devcuration.com/articles/relay-raises-36m-frontline-operational-intelligence

Miter

On a construction job, 1 hour can belong to a worker, a union, a cost code, a prevailing-wage rule, a state tax regime and a project budget.

The crew experiences an hour. The back office inherits the paperwork.

That is the operating problem behind Miter's $40M Series B. Battery Ventures led the round, with existing investors Bessemer Venture Partners and Coatue Management participating. Miter says total capital raised is now $78M.

Connor Watumull and Tobin Paxton did not find this in a market map. Watumull worked inside a construction bookkeeping firm. Paxton watched his mother wrestle with payroll on Thanksgiving. They spent time with Bay Area contractors, saw how much judgment was trapped between the field and finance, and started Miter in 2021.

The product began with construction payroll and workforce management. It now connects HR, compliance, expenses, time tracking, safety, field operations, accounts payable, and job costing. Miter says more than 2,000 contractors and 150,000 users are on the platform, with more than $8.5B in payroll processed each year.

Those are company-reported numbers. The business meaning sits in the connection between them. A time entry changes payroll. Payroll changes taxes, fringes, and certified reports. The same hour also changes the labor cost of a project that may already be drifting from budget.

Generic business software can hand those obligations to separate systems and call the integration somebody else's problem. Construction cannot. Crews move, pay rates vary, unions have their own rules, public projects add prevailing-wage reporting, and every mistake eventually reaches a worker, a regulator, or a margin.

Miter is using the new capital to expand AI products across the back office and jobsite, grow engineering in San Francisco and New York, and add U.S. sales teams. The company says its AI already structures data from invoices, receipts, and pay tables, helps automate manual workflows, surfaces labor and job-cost trends, and creates safety reports from photos or voice notes.

Battery is underwriting the expansion from payroll software toward a construction operating system. Bessemer and Coatue are following their earlier investment after Miter says customer count tripled in 18 months.

That expansion raises the standard with the opportunity. When software sits between a foreman's timecard and a worker's paycheck, an AI suggestion is not a clever demo. It becomes part of the record the company has to defend while the job keeps moving.

Deeper Analysis 👇

https://devcuration.com/articles/miter-raises-40m-construction-operating-system

Skylight

Consumer hardware makes a company pay for success before it gets to recognize the revenue.

Skylight can sell a family on a calendar, photo frame, or kids' routine device in a few clicks. It still has to forecast demand, commit to inventory, move products through channels, and have the right box available when a household is done with the kitchen whiteboard. That timing gap is where this financing lives.

Skylight has added $25M to its credit facility with SG Credit Partners and Wingspire Capital, increasing total capacity from $50M to $75M. This is debt, not a venture round, and $75M is the full facility size rather than $75M of newly raised cash.

Skylight says it has remained bootstrapped while building a physical-product business at consumer scale. It reports that the business doubled over the past year and now reaches more than 12M customers, including 1.7M households using Skylight Calendar. Those figures explain why working capital has become part of the product roadmap.

Michael Segal founded Skylight around a simple idea: make digital family information useful in a physical place. Jake Kring, CTO, helped turn that idea from a photo frame into a broader technology platform. Chia Chung, COO and CFO, is now managing the less photogenic part of that expansion: enough liquidity to fund inventory, new devices, and additional distribution without selling equity to do it.

The facility has grown with the operating model. SG Credit Partners announced $15M of senior debt in 2023. By April 2025, SG said a prior $20M facility had expanded to a $50M loan, with Wingspire joining the partnership. The latest $25M increase takes the line to $75M.

The new capacity is meant to support more products, channels, and markets. That includes continued development of Skylight Buddy, the children's routine device the company says sold out within a week, and deeper investment in Skylight Assistant, a text-based service designed to coordinate schedules, reminders, and household tasks.

Skylight is making a specific bet about family technology: the winning interface may be the one that becomes part of the house instead of asking every person in it to open another app. That bet creates a supply-chain obligation alongside the software ambition.

Debt can preserve ownership, but it also makes forecasting discipline visible. The next phase will be decided in purchase orders, retail shelves, renewal behavior, and whether each new product earns a place in routines already crowded with devices.

Deeper Analysis 👇 https://devcuration.com/articles/skylight-adds-25m-expands-credit-facility-to-75m

Proximal

A useful model failure is becoming one of the most valuable raw materials in AI.

The easy version of the data business was volume: collect more examples, hire more labelers, and give the model a larger pile of human judgment. That model weakens as frontier systems outperform more people in narrow technical domains. The harder question is how to discover what a model still cannot do and turn the failure into the next lesson.

Proximal is building around that loop. Co-founders Calvin Chen, CEO, and Justus Mattern have announced a $15M Seed led by General Catalyst at a $300M valuation. SV Angel, Diede van Lamoen of Go Global Ventures, Chemistry, Liam Fedus, Kevin Weil, and Erik Bernhardsson are also among the backers.

The timing deserves precision. Proximal made the financing public on September 29, but Calvin Chen told The Information the round closed at the end of 2025, when revenue was negligible. This is a new announcement of an older close, not $15M that appeared in the bank this week.

What changed between the close and the announcement is more interesting than the calendar. Proximal says it has reached $200M in annualized revenue while working with frontier labs and enterprises on coding agents. The Information says the figure annualizes current-quarter revenue. It is a company-reported run rate, not audited annual revenue, and Proximal has not named its customers or disclosed concentration.

The operating argument is unusually clear. Proximal turns raw agent traces and artifacts from real workflows into evaluations that expose capability gaps. Those signals guide targeted post-training data and model deployment. Software engineering is the proving ground because code can be executed, tested, broken, restored, and measured with unusual precision.

Proximal has made some of that work public through FrontierSWE. Its second version puts models through 34 ultra-long-horizon engineering and research tasks with 20-hour agent budgets. The benchmark is not a revenue audit. It shows that Proximal understands the expensive part of the category: designing environments where failure teaches something useful instead of producing another leaderboard number.

Proximal is hiring researchers and engineers in San Francisco and Bangalore while it builds that expansion. The company is trying to make model weakness observable, programmable, and useful. If that loop holds across domains, the next scarce dataset may begin with a failure somebody finally learned how to ask for.

Deeper Analysis 👇 https://devcuration.com/articles/proximal-raises-15m-to-turn-ai-failures-into-training-data

R1 Therapeutics, Inc.

Phosphate is easy to ignore until damaged kidneys can no longer clear enough of it and dialysis patients inherit another daily management problem.

R1 Therapeutics is financing a different way to attack that problem. The company expanded its Series A from $77.5M to $80.5M after Samsara BioCapital joined the syndicate, adding fresh capital while R1's lead program, AP306, is already inside a global Phase 2b trial.

The amount moved by $3M. The evidence burden did not move an inch.

AP306 is an investigational pan phosphate transporter inhibitor designed to block 3 active transport pathways in the gut. Traditional phosphate binders have carried much of this job for decades, yet R1 says more than 70% of U.S. dialysis patients still fail to reach normal phosphate levels. The company is studying AP306 as a monotherapy, with the possibility of lowering phosphate without recreating the same treatment burden in a different bottle.

That possibility now belongs to a randomized, double-blind, placebo-controlled study. R1 and Alebund Pharmaceuticals plan to enroll roughly 168 adults receiving maintenance hemodialysis across the U.S. and China. Six fixed-dose AP306 regimens will be compared with placebo over 8 weeks, and topline data are expected in H1 2027.

Krishna Polu, M.D., co-founder, president and CEO, and L Mary Smith, Ph.D., co-founder and COO, have built R1 around a single clinical-stage asset licensed from Alebund outside Greater China. That concentration creates clarity and pressure in equal measure. The company does not have a portfolio to hide behind if the mechanism fails to carry into a larger study.

The investor group understands both sides of that pressure. Abingworth, DaVita Venture Group and F-Prime co-led the original Series A. Curie.Bio, SNMA Capital and U.S. Renal Care also participated. Now Abraham (Abe) Bassan and Samsara are entering after the trial started but before the data arrive.

This is where biotech capital stops being an applause meter and becomes operating time. The financing can support sites, dosing, monitoring, manufacturing and the unglamorous discipline required to produce a clean readout. It cannot rescue an endpoint or negotiate with blinded data.

R1 has a differentiated mechanism, published Phase 2a evidence and kidney-care investors around the table. By H1 2027, the conversation will belong to the patients, dose arms and serum-phosphate measurements moving through NCT06712654.

Deeper Analysis 👇

https://devcuration.com/articles/r1-therapeutics-expands-series-a-to-80-5m

Ascerta

A board can approve an AI budget while every dashboard underneath it reports a different version of success.

Engineering counts tokens, model calls, and agent runs. Finance sees cloud commitments and licenses. Then somebody asks which initiative created value, and activity has been doing a convincing impression of an answer.

That gap is the business behind Ascerta. The Bellevue company, formerly known as Pay-i, has raised an $18M Series A led by Dell Technologies Capital, with Hitachi Ventures, BGV, Wipro Ventures, and earlier investors participating. Total funding is now $22.9M.

Co-founders David Tepper, CEO, Doron Holan, CTO, and Erik Winters, COO, started with AI cost management. Customers pulled the problem outward. They wanted to know how people used AI, what agents did, and whether the work justified more investment.

The rebrand follows that mandate. Ascerta calls the category Enterprise AI Management, ambitious language for a practical problem: the people buying, building, using, and budgeting AI rarely work from the same evidence.

Atlas measures adoption, value, ROI, and waste. Forge examines how coding-agent adoption becomes productivity. Convoy helps companies manage provisioned AI capacity without leaving expensive commitments stranded.

The platform follows activity by person, team, tool, and use case, then ties it to the business outcome it was supposed to change. A model call becomes useful when finance can trace it to a workflow, engineering cycle, or capacity decision somebody owns.

Ascerta says customers have improved AI-initiative ROI by 47%, shortened agent-launch time by 24%, and cut wasted AI spend by 86% on average. Those are company-reported results, not an independent audit. They describe what buyers expect: fewer vanity metrics and a defensible answer about the next dollar.

Dell Technologies Capital is underwriting that system-of-record thesis. Wipro is both investor and customer. The strategic possibility is that the measurement layer travels with infrastructure, integrators, and operating teams shaping enterprise AI deployments.

The $18M will expand Ascerta's platform, engineering, go-to-market work, and integrations. Once a tool helps decide which AI programs get scaled, fixed, or cut, its measurements enter the argument over capital, jobs, and operating priority.

Enterprise AI has spent years proving it can generate activity. Ascerta is building for the meeting where that activity has to defend its budget.

Deeper Analysis 👇 https://devcuration.com/articles/ascerta-raises-18m-prove-ai-spend-pays-off

Beltic

An AI agent arrives at a payment endpoint wearing somebody else's credentials and carrying instructions that may have passed through 3 other agents.

The merchant gets a request. What it needs is a chain of authority.

Beltic raised a $7.3M Seed led by Norwest, with Restive Ventures, Oxford Seed Fund, and Collide Capital participating. Total disclosed funding is now $8.8M after an earlier $1.5M Pre-Seed led by Collide Capital.

Isha Bhatnagar and Mike Allan built Beltic around a problem their careers kept putting in front of them. Bhatnagar, co-founder and CEO, spent 6 years at Coinbase working on identity, onboarding, and global regulatory experience. Allan, co-founder and COO, built Brazilian banking-as-a-service company Atar through its 2021 acquisition by Porto Seguro.

CTO Farhan Afsahi adds the data layer. Afsahi co-founded Verifiet, which Beltic acquired in 2025. Beltic and Norwest say the infrastructure connects to global jurisdictions and covers more than 500M entities. That scale is company and investor reported. The product logic is still clear: an agent credential is only as credible as the verification underneath it.

Beltic calls the category Know Your Agent. The product combines visibility, policy, and intent. It identifies the agent and operator, checks the action against policy, and evaluates intent across websites, APIs, MCP servers, and payment rails.

That distinction will matter as agentic commerce leaves the demo and reaches systems where a bad decision can repeat thousands of times before anyone notices. A valid login does not explain whether software was authorized to transfer $1,000, call another agent, change the recipient, or keep going after the original task changed.

The market is already building transaction rails. Visa has a Trusted Agent Protocol. OpenAI and Stripe have the Agentic Commerce Protocol. Beltic is placing its bet one layer closer to the decision: who delegated the authority, what survived each handoff, and whether the action should be allowed now.

The company says it is working with select design partners, but it has not named them or published audited adoption metrics. The $7.3M buys time to prove that an integrated verification stack can stay fast enough for autonomous transactions and defensible enough for the risk teams asked to approve them.

Agentic commerce will move at machine speed. Beltic's opportunity lives in keeping identity, authority, and accountability connected while the machines keep handing work to one another.

Deeper Analysis 👇 https://devcuration.com/articles/beltic-raises-7-3m-to-verify-ai-agent-transactions

CScale

A failed laser sounds like a component problem until thousands of accelerators are waiting on the connection around it.

That pressure sits behind CScale's $145M Series C. The company exited stealth with optical interconnect for AI scale-up, where processors across dozens of racks must communicate like one machine.

Atreides Management, Valor Equity Partners, and Premji Invest co-led the round. Sutter Hill Ventures and Maverick Silicon continued their support, while NVIDIA and Intel Capital joined as CScale's first strategic investors. Total reported funding is now $188M.

CScale is building an integrated light engine around a specific assumption: at fleet scale, rare failures stop behaving like rare events. Its architecture is designed to keep the surrounding compute available when one laser fails.

That changes the commercial argument for optics. Bandwidth and reach get the headlines, but utilization pays the infrastructure bill. Keeping expensive accelerators useful through an optical failure may separate a component event from a capacity event.

Founder and CTO Sanjai Kohli has spent a career turning hard communications technology into products, including helping build SiRF around mass-market GPS. CEO Martin Lund came from leading Cisco's Common Hardware Group across silicon, systems, and optics. The handoff now is from architectural conviction to a product that customers can qualify, deploy, and trust.

The investor list fits that work. NVIDIA and Intel Capital sit close to the compute ecosystem CScale wants to serve, while the co-leads are funding the path from stealth engineering to commercial infrastructure.

The evidence has boundaries. CScale disclosed no valuation, named customer, production deployment, or benchmark. Martin Lund told Reuters the company plans to ship chips by 2028, so the round is financing the build, not celebrating a finished system.

That is what makes the story useful. AI infrastructure is becoming less forgiving as clusters absorb more processors, more power, and more capital. The interconnect cannot merely carry the traffic. It has to preserve the value of the compute when failure becomes part of ordinary fleet behavior.

CScale now has $145M to turn that reliability thesis into something a data-center operator can measure under load. Its commercial value will take shape when a failed light source becomes an ordinary contained event and the surrounding machines keep working.

Deeper Analysis 👇 https://devcuration.com/articles/cscale-raises-145m-for-resilient-ai-optical-interconnect

Kanu AI

When a senior employee knows which spreadsheet matters, which exception needs judgment and which approval cannot be automated, the company owns valuable logic without owning the software that carries it.

Kanu AI has raised a $11.7M Seed to change that arrangement. Trilogy Equity Partners led the round, with a16z speedrun, BMW i Ventures and Accel participating.

Kanu asks employees to show the system how work gets done. It connects the relevant data and tools, turns that demonstrated process into operational software, and deploys the result inside the customer's cloud. Infrastructure, permissions, model accounts and data boundaries remain under customer control.

Enterprise work is full of decisions that never made it into a product requirements document. They live in inboxes, spreadsheets, CRM notes and the heads of people who can spot the exception before it becomes an expensive mistake.

Karan Grover, CEO and co-founder, and Harsh Patel, COO and co-founder, are building Kanu around the idea that this judgment can become an asset the organization owns. Kanu says one employee can teach the platform a better workflow and make that improvement available across the company without another long software rollout.

Kanu's answer is governance close to the work: deployment inside AWS, Google Cloud or a private environment, inspectable execution, model choice under customer control and human approval where judgment still belongs to a person.

The company reports that revenue has more than doubled in every quarter since launch, with customers across commercial real estate, insurance, financial services and technology. For one unnamed national customer, Kanu says an analysis process that took as long as 8 weeks now runs in under 10 minutes, with expected annual software-cost savings above $1M and millions in projected additional revenue.

Those results are company-reported, and the customer is not named. They still make the economic idea legible. The prize is not another chat window. It is the operating knowledge that companies already paid employees to develop but have never been able to compound cleanly.

That transition will be measured each time a customer changes a rule, adds a data source or asks the system to explain why it acted. Kanu is financing the machinery that turns those changes into software across teams without letting the company's own judgment disappear inside the machine.

Deeper Analysis 👇 https://devcuration.com/articles/kanu-ai-raises-11-7m-for-enterprise-workflow-software