❯ Amazon in talks to buy AI startup Decart for about $7 billion, which could give its in-house chips a software layer

Could close within daysAmazon is in talks to acquire AI startup Decart in a deal that could be worth about $7 billion, The Wall Street Journal reported on October 11, citing people familiar with the matter. According to SiliconANGLE’s account of the report, a deal could be completed in the coming days if talks do not break down and no other bidder emerges.

Software for switching chipsDecart was founded in 2023 by three Israeli engineers. Its core product is software called DOS, which helps developers tune AI models to run efficiently on different chips. Optimizing a model for a specific chip used to take months and can now take about two weeks, making it easier for companies to change platforms when chip supply tightens or needs shift. The company also builds world models such as Lucy and Oasis, which generate realistic 3D environments for games and robot training.

Anthropic walked away last monthDecart raised $300 million in May at a $4 billion valuation from investors including Radical Ventures, Nvidia Ventures, Adobe Ventures, Sequoia and Benchmark. Bloomberg previously reported that Anthropic was in talks in September to buy Decart for about $6 billion and pulled out after due diligence, for reasons that are unclear.

The missing piece for in-house chipsAmazon Web Services sells its own AI training chip, Trainium3. Running a model on a different chip means optimizing it again, which is the problem Decart’s software solves. The report said Decart received early access to Trainium3 last year. If the acquisition goes through, Amazon could hand the tool directly to customers and lower the barrier to moving onto its chips.

▮ SIGNALBuilding a custom chip is only the first step; whether customers move depends on how painful migration is, and cloud providers are now paying heavily for the moving tools.

❯ Nvidia in talks to acquire or invest more in open-weight model maker Reflection AI, moving the chipmaker further into models

Full purchase or just the teamNvidia is in early talks with US AI startup Reflection AI, the Financial Times reported on October 10, citing people familiar with the matter. Options include a full acquisition, hiring the team and licensing its technology, or adding investment and compute. According to TIKR’s account, a deal could come within weeks or fall apart; Reflection declined to comment.

Models you can download and run yourselfReflection builds open-weight models, which companies can download, modify to their needs and run on their own hardware. On October 5 it released its first open-weight model, Beam, aimed at coding and agentic tasks. Nvidia has its own family of open models called Nemotron, and buying Reflection would expand its position in the model layer.

$800 million already investedNvidia has already invested $800 million in Reflection. In April, Reflection CEO Misha Laskin told CNBC the company was raising at a $25 billion pre-money valuation, the most recent publicly known price. No acquisition price for the current talks has been reported.

What "hiring only" avoidsHiring staff and licensing technology can skip the lengthy regulatory review a full acquisition usually faces, according to PYMNTS. Nvidia used a similar approach with chip company Groq last December, licensing its technology and hiring its CEO in a deal CNBC valued at about $20 billion. Which option is chosen determines whether Reflection’s other shareholders receive acquisition proceeds or are left with a company whose core team is gone.

▮ SIGNALThe better open models get, the more companies want to buy chips and run them in-house, which gives a chip seller reason to nurture models itself.

❯ Nvidia plans to invest in inference chip rival d-Matrix, bringing a competitor’s chips into its own racks

Three people with knowledgeNvidia plans to invest in AI inference chip company d-Matrix, The Information reported on October 10, citing three people with knowledge of the deal. According to tbreak, the size and timing were not disclosed and neither company has confirmed it. d-Matrix’s chips are used for inference, the stage where a model responds to user requests, and compete with Nvidia’s products.

A rival's chip in an Nvidia rackThe two companies already work together. d-Matrix said on September 10 that it will connect its next-generation Raptor processors to Nvidia’s platform through NVLink Fusion. NVLink Fusion is an interconnect technology Nvidia introduced in May 2025 that lets other companies’ accelerators plug into Nvidia’s racks and networking, so data centers can mix processor types. Marvell reached a similar partnership with Nvidia in March and received an investment from it.

Valuation and the wider portfoliod-Matrix’s last round was a $275 million Series C in November 2025 at a $2 billion valuation; The Information reported in July that it was seeking new financing at $5 billion. A separate report from The Information said Nvidia is also scouting deals in robotics, self-driving cars and on-device models, and held nearly $100 billion in equity investments as of late July.

Customers switch chips, the business staysThe Information noted that customers are seeking alternatives to Nvidia’s GPUs. If those alternative chips run inside Nvidia’s racks and networking, the order for the overall infrastructure stays with Nvidia even when customers buy fewer GPUs. A challenger like d-Matrix gets into data centers faster by joining an existing ecosystem, at the cost of being tied more closely to its largest rival.

▮ SIGNALNvidia’s answer to alternative chips is to absorb rivals into its own system; what it is defending now is the standard for the whole rack.

❯ Anthropic’s CEO asked Meta for compute and was turned down, as the shortage pushes rivals to ask each other for help

A call in search of computeAnthropic CEO Dario Amodei contacted Meta chief AI officer Alexandr Wang earlier this year hoping to secure more compute, and Meta did not agree, The Wall Street Journal reported, citing people familiar with the matter. The report’s standfirst, as listed on Techmeme, says bitter rivals are forming alliances and executives are having to step in personally to secure resources.

A $10 billion lease discussed in JulyAI News Weekly noted that CNBC reported in July that Anthropic had proposed leasing compute from Meta, with preliminary talks covering about $10 billion over two years. The Journal did not say whether Amodei’s outreach was part of the same talks, how much compute he sought or why Meta declined.

What is scarce is chips usable todayTraining new models and serving user requests both consume chips, and how much compute a model company can get directly determines how many users it can serve. For Meta, handing compute to a directly competing model company amounts to speeding up a rival, which makes the deal hard to settle on price alone. Another Journal report listed on the same Techmeme page said Anthropic co-founder Tom Brown brokered a compute deal with SpaceX worth $1.25 billion a month.

▮ SIGNALAfter a model company raises the money, someone still has to be willing to give up ready compute, and those who have it are often its competitors.

❯ Anthropic books gross cloud sales while OpenAI books only its share, so their annualized revenue figures cannot be compared directly

One metric, two methodsAnthropic and OpenAI calculate annualized revenue differently, Bloomberg reported on October 9. For sales made through cloud partners such as Amazon, Anthropic counts the gross amount as its revenue, while OpenAI recognizes only its own share. The report said Anthropic’s annualized revenue reached $65 billion in July and OpenAI expects to reach or exceed $70 billion by the end of this year.

One dollar sold: one books a dollar, the other 20 centsTechTimes gave an example: when Microsoft Azure sells $1 of OpenAI model usage, OpenAI records only its roughly 20-cent share as revenue; in the same situation Anthropic records the full $1 and lists the cloud provider’s cut as a cost of sales. Annualized revenue itself is only a full-year figure extrapolated from recent pace.

Between $50 billion and $70 billionThis explains a set of numbers that recently confused investors. Citing a Financial Times report, the article said OpenAI’s annualized revenue was approaching $50 billion at the end of September, about $20 billion below the $70 billion figure that had circulated, and attributed the gap to different accounting and an investor calculation error. Neither company is public, so outsiders compare them using scattered disclosures, and when methods differ there is no definite answer to who has “more revenue.”

▮ SIGNALBoth companies are heading toward public listings, and prospectuses will put revenue methods on the table; the rankings now in circulation may be reshuffled.

❯ OpenAI raises GPT-6.1 Sol’s default context to 500,000 tokens, with no long-context fee for subscribers

Default raised by about 80%OpenAI’s Thibault Sottiaux announced on X on October 11 that the default context length for GPT-6.1 Sol, when used through Codex and ChatGPT Work, has gone from 272,000 tokens to 500,000 tokens, an increase of about 80%. Users can still set it higher or lower, up to a maximum of 1 million tokens. He also said long-context fees are not charged when usage is drawn from a subscription.

Fewer compactions, fewer lost detailsContext length determines how much a model can “remember” at once. On long tasks, when the conversation and code exceed the limit, the tool compacts earlier content into a summary and details can be lost. Sottiaux said the higher default reduces compactions and helps with tasks that need to keep a lot of context and nuance over time. He said the value was carefully tested, is the sweet spot for most users and will go higher in the future.

Day 7 of a 28-day pledgeThis is day 7 of an ongoing OpenAI commitment. According to a weekly roundup by X user @btibor91, OpenAI has pledged to ship one improvement to Codex and ChatGPT Work every day for 28 days, or give users a full usage reset on any day without one. Sottiaux also ran a poll on Pro plan allowances and said afterward he had expected users to favor 2x usage in exchange for bringing back 5-hour limits, and that the result confused him.

▮ SIGNALCompetition among coding tools has come down to how much code fits in one go and how allowances are counted; shipping an improvement a day has itself become a way to keep heavy users.

❯ An unreleased life sciences model, gpt-rosalind-discovery, appears on OpenAI’s pricing page at the same price as the existing one

A new line on the pricing pageOpenAI’s API pricing page has added gpt-rosalind-discovery under “Life sciences models,” next to the existing gpt-rosalind-research, TestingCatalog reported on October 10. The two are priced the same: $5 per million input tokens and $25 per million output tokens. OpenAI has published no announcement, help article or system card for it.

Sold only to vetted organizationsGPT-Rosalind is the life sciences model OpenAI introduced in April for biology, drug discovery and translational medicine, with early partners including Amgen, Moderna, the Allen Institute and Thermo Fisher. On September 11 it left research preview and opened to eligible organizations worldwide, though they must be approved through a trusted-access program. Its system card says the model is trained not to refuse sophisticated biology questions, so gated access is the main safeguard.

The name points to drug discovery; the purpose is unconfirmedThe name suggests the new model may target drug discovery, such as target identification or molecular design, but TestingCatalog said this is a guess. OpenAI has not said whether it is a separately trained model or a differently tuned version of the existing one. OpenAI’s help center says the current model is most useful for early discovery and hypothesis generation, such as literature synthesis and omics interpretation; what the new one adds will have to wait for an official description and testing.

▮ SIGNALGeneral-purpose model companies are starting to split out industry-specific models and to control risk through gated access instead of the model’s own refusals.

❯ Microsoft releases decision-scoring model Microsoft-Decision-1, built on an Alibaba open-weight model to answer multiple-choice questions for agents

It scores, it does not writeMicrosoft released Microsoft-Decision-1 on October 9, announced by CEO Satya Nadella on X. According to TestingCatalog, it does not generate open-ended text. It takes a fixed set of options and returns a probability for each, supporting yes/no, multiple-choice and rating questions, and can grade AI responses and agent actions against a rubric. The model is available in Microsoft Foundry at $0.042 per million input tokens, with output tokens free.

The base comes from Alibaba's QwenThe model was post-trained from Qwen3.5-9B, Alibaba’s open-weight model with 9 billion parameters. Microsoft said it will later release versions based on its own MAI models and OpenAI models, without giving a timeline. RuntimeWire said the introductory post was written by Microsoft vice president Achint Srivastava, who previously co-founded AI evaluation company Pi Labs.

35x faster than GPT-6 Sol in Microsoft's own testsMicrosoft said the model had the highest accuracy across 36 benchmarks covering nearly 150,000 questions and was 35 times faster than GPT-6 Sol at median latency. These are Microsoft’s own results and have not been verified by third parties. Microsoft also said Xbox Research used it to categorize more than 10,000 pieces of player feedback with quality comparable to GPT-6 Sol at 200 times lower cost.

A traffic light for agentsMicrosoft positions it as a low-cost control layer in agent systems. At every step an agent has to decide whether to continue, retry, switch tools or hand off to a person. Having a large model make those calls is slow and expensive; with a dedicated small model, developers can add a check at every step without worrying about the bill. It handles only text and fixed options and cannot replace the large models that do the reasoning and generation.

▮ SIGNALThe model that runs most often in an agent system is not necessarily the strongest one; routing, review and scoring are being handed to cheap, specialized small models.

❯ Apple to unveil smart home hub HomeView on Tuesday, with accessories made by LG, betting on a voice-controlled home

A home hub with a screenApple will unveil a smart home hub called HomeView on October 13, according to Bloomberg’s Mark Gurman. TechCrunch described it as a Siri-powered device with a display and an A18 chip that controls other smart devices in the home, in countertop and wall-mounted versions. A new Apple TV and HomePod mini will be unveiled alongside it.

No App StoreGurman added details on X: HomeView will not have an App Store and will run a limited set of Apple apps, centered on Siri and connected accessories; the new Apple TV moves to the A19 chip and the HomePod mini to the S10. He expects HomeView to ship around mid-November, while the Apple TV and HomePod mini arrive on October 30.

Accessories under the LG brandAccessories such as a lock, thermostat and doorbell were developed by Apple with LG. According to Macworld’s account of Gurman’s reporting, they will not carry Apple’s logo, will be sold under the LG brand, and LG will handle manufacturing and support, an arrangement Gurman calls “second-party.” TechCrunch noted that Apple’s smart home hub was previously delayed by problems with Siri.

Why View and not PadGurman explained that Apple wants users to look at the display and control it mainly by voice; it supports touch, but “Pad” would suggest an iPad-like experience. How well HomeView works therefore depends largely on whether the new Siri can understand requests and get them done. TechCrunch called it Apple’s most serious smart home bet to date, after HomeKit and HomePod failed to gain much traction.

▮ SIGNALApple is moving the entry point to the smart home from a phone app to an always-on screen and a voice assistant; for the first time, how smart the assistant is directly decides a hardware product’s fate.

❯ Tim Cook meets Guangdong Governor Meng Fanli in Guangzhou, says Apple is willing to expand its business in the province

A meeting in Guangzhou on October 11On October 11, Meng Fanli, deputy Party secretary and governor of Guangdong province, met Apple executive chairman Tim Cook and his delegation in Guangzhou, with provincial official Zhang Guozhi attending, according to the Nanfang Plus news app. 21jingji.com carried the report.

Supply chain, manufacturing and channelsCook thanked Guangdong for its long-term support of Apple’s development and said the province is a major global manufacturing center with complete industrial and supply chains, strong innovation and a consumer market with great potential. He said Apple is willing to further expand its business in Guangdong and to explore cooperation with relevant cities and companies in supply chain coordination, the innovation ecosystem, smart and green manufacturing and sales channel expansion.

An invitation to place R&D platforms in GuangdongMeng welcomed Cook and his delegation. He said Apple and Guangdong have long cooperated in production, sales and research and development, and that Apple is welcome to place more innovative products, emerging businesses and R&D platforms in the province. Guangdong will expand high-level opening up, build a first-class business environment and provide services to all kinds of companies including Apple. The official readout did not announce specific projects or investment amounts.

▮ SIGNALWhen Apple talks cooperation with Guangdong, supply chain and manufacturing come first; for Apple this is above all the place where its products get made.

❯ ByteDance spin-off Anew Labs raises $290 million in an angel round at a first-round valuation of RMB 10 billion

RMB 10 billion in the first roundAnew Labs, an AI drug discovery company spun out of ByteDance, completed a $290 million angel round in September at a first-round valuation of RMB 10 billion, The Economic Observer reported on October 10. Investors include Sequoia China, IDG Capital, Hillhouse Ventures, 5Y Capital, Gaorong Ventures, Primavera Venture Partners and Boyu Capital, with Sino Biopharmaceutical and the Shanghai Future Industry Fund participating as strategic investors.

Building models and its own drugsThe team was formed inside ByteDance in 2021 under Liu Kai, became independent in June this year, is headquartered in Shanghai and has fewer than 100 people. It develops models, including the open-source protein models Protenix and PXDesign and AnewMind for drug R&D decision-making, while also advancing its own drug pipeline. ByteDance remains the controlling shareholder after the spin-off and continues to supply compute through Volcano Engine.

All four programs are preclinicalThe company has four preclinical pipeline assets. The lead program is an oral small-molecule candidate targeting IL-17, now in lead optimization. The company says it is the world’s first full-spectrum oral small-molecule inhibitor targeting both IL-17A and IL-17F, with data first presented in April at the annual meeting of the American Association of Immunologists. Preclinical means the drug has not yet entered human trials.

People who know algorithms and people who know drugsIndustry figures quoted in the report said returns from tools alone are limited and a pipeline offers more room, but a pipeline needs senior pharmaceutical talent to oversee it. Internet companies are used to fast iteration, while drug development is long and failure-prone; how the two kinds of talent work together and how returns are measured are problems the company has to solve. Its scientific advisers already include people from drugmakers, such as former Innovent Biologics president Liu Yongjun.

▮ SIGNALWhen an internet company goes into AI drug discovery, compute and algorithms are the advantages it brings; whether it produces a drug depends on rebuilding a team by the rules of life sciences.

❯ Neolix’s driverless fleet reaches 27,000 vehicles as Shenzhen night deliveries expand to 437 routes

A driverless van at 10 p.m.Beijing-based Neolix operates about 27,000 driverless delivery vehicles across 300 cities in 15 countries, the world’s largest autonomous fleet, Bloomberg reported. At a farmers’ market in Shenzhen’s Longgang district, a grocery manager summons a van by phone after 10 p.m. and loads it with poultry bound for customers across town.

Shenzhen opened night delivery in MarchShenzhen opened nighttime deliveries to driverless vehicles in March, installing anti-glare lighting and setting new vehicle standards and an emergency response protocol under a trial regulatory framework. The trial runs only on approved corridors, and vehicles must follow pre-designated routes and areas. Since it began, Neolix’s overnight service has expanded to 437 routes. The company is testing overnight logistics with SF Express, moving parcels at Futian Station from an underground staging area to freight trains.

Empty roads at night, but harder drivingRoads are emptier and less congested at night, idle vehicles and infrastructure can work another shift, and no night-shift wages are needed. The difficulties are just as clear: cameras perform worse in low light, and lidar can struggle with headlight glare and reflections off wet surfaces. The hardest part is the final leg, with loading docks, unmarked spots and entrances missing from maps; fewer workers are on site at night, so vehicles must be more precise.

A target of 50,000 this yearNeolix CEO Yu Enyuan said the company expects its domestic fleet to reach 50,000 vehicles this year and plans to add 100,000 more next year. He said night delivery lowers costs and reduces the physical strain on couriers. With an extra overnight shift of capacity, couriers and fresh-food merchants can have next-morning goods in place earlier.

▮ SIGNALDriverless delivery vans reached scale first in the most constrained settings, low speed, fixed routes and night hours, a more pragmatic path to commercializing autonomy than carrying passengers.

❯ DJI and Insta360 together ship more than nine in ten handheld smart cameras worldwide as GoPro’s share falls to 3%

Two Shenzhen companies take 93%In the second quarter, DJI and Insta360 held 73% and 20% of the global handheld smart camera market, while action camera pioneer GoPro’s share fell to 3%, the Financial Times reported, citing IDC data. The two Shenzhen companies together hold 93% and are competing for the global top spot.

5.88 million units in a quarterThe IDC report, as summarized by RedShark News, shows global handheld smart camera shipments of 5.88 million units in the second quarter, up 57% year over year, with sales of $2.34 billion. DJI shipped 4.28 million units, up 87%; Insta360 shipped 1.15 million, up 40%, and holds 73% of the 360 camera segment. GoPro’s action camera shipments fell 49% and its 360 camera shipments fell 61%.

Growth comes from pocket gimbal camerasThe category covers action cameras, 360 cameras and gimbal cameras. The fastest growth came from gimbal cameras such as DJI’s Osmo Pocket, at 2.8 million units, up 87%, overtaking traditional action cameras as the largest segment. Over the same period the average selling price fell 6% for gimbal cameras and 4% for action cameras. IDC expects shipments to approach 41 million units by 2030.

Meeting head-on in the same categoriesThe two companies’ product lines are overlapping. IDC data compiled by PhotographyTalk shows that in wide-angle action cameras DJI holds 64%, up 11 points from a year earlier, and Insta360 more than 17%; shipments of detachable thumb cameras grew more than 300%, with Insta360 holding 43%. With both competing directly in the same categories, new products and price changes will come more frequently.

▮ SIGNALGrowth in this market is driven by new categories; gimbal and thumb cameras have rewritten the rankings, and companies that stuck to a single category fell behind fastest.

❯ Karpathy says the gap between how the public and heavy users see AI is widening, with most people having only met a “derpy chatbot”

A funnel that keeps narrowingAI researcher Andrej Karpathy posted on X on October 11 that the gap in shared understanding of what large language models can do is widening, in the shape of a sharp funnel. By his rough estimate, about 6 billion people have barely come into contact with LLMs, and 1 to 2 billion use free-tier chat products occasionally, treating them as a better search engine or writing aid. The post has passed 1.55 million views.

20 million people and 5,000 peopleFurther in are about 20 million professionals using frontier models for math and code, who see first-hand that large projects that once took weeks or months can now be completed by agents from a prompt. At the innermost layer are about 5,000 people with internal access to the most advanced systems. He describes swarms of thousands of agents collaborating over weeks on large software projects, cyber attack and defense, and frontier mathematics, with human review and comprehension starting to fall behind.

Three reasonsKarpathy attributes the divide to three factors. First, impact scales with the size of the problem; a question answered in a paragraph barely uses the system’s capability. Second, capability is uneven: models are strongest in domains that are digital, verifiable and economically valuable, because their ability comes from reinforcement learning on verifiable rewards. Third, access: free, paid and internal tiers see different things.

Three experiences at the same momentHe summed up that the general public sees a derpy chatbot, most professionals feel only a modest uplift, and a small sliver are experiencing the vertigo of a curve going vertical, all at once. That explains why discussions about AI so often talk past each other: the two sides may be using products several tiers apart. Managers who judge only by their own experience with a free tier may underestimate what is changing.

▮ SIGNALAI capability spreads unevenly; to judge where the technology stands, first ask which tier of product the speaker uses and how big a problem they point it at.

❯ Yuandong Tian says his research work may be replaced by AI in about five years, so he started a self-improving AI company

"Better than losing the job passively"Former Meta FAIR researcher Yuandong Tian said on the podcast The Information Bottleneck that his work will probably be replaced in about five years, and that it is better to take the initiative and start a company in this direction than to lose his job passively. According to a write-up by Synced republished by 36Kr on October 10, his last paper at Meta studied grokking, in which a neural network first memorizes and later generalizes; it was single-authored, and he said he regarded GPT-5 as a co-author.

From Go AI to self-improvementTian went to Carnegie Mellon University for his PhD in 2008, spent a little over a year on Google’s self-driving team after graduating, and then worked at Meta FAIR for nearly a decade, where he built the Go programs DarkForest and OpenGo. He is now a co-founder of Recursive Superintelligence, which researches AI that can improve itself. In May it announced more than $650 million in funding at a $4.65 billion valuation, led by GV and Greycroft with Nvidia and AMD participating.

AI's own ideas are still mediocreHe said on the show that when a model proposes research ideas on its own, it tends to produce mediocre directions almost everyone already knows. As models get stronger, though, researchers may need to step back and think at a higher level instead of writing hypotheses one by one. He also said engineering skill has become more important, and AI coding tools let experienced researchers test ideas and find the key problems quickly.

Why latent reasoning has not been adoptedCoconut, a method he co-developed, lets a model reason in a continuous latent space instead of generating text token by token, an idea that came from his observation that he does not use language when he thinks. He gave two reasons frontier models have not adopted it: there is little training data recording “latent thought,” and such internal representations may differ from model to model. On reinforcement learning, his takeaway was that how the action and state spaces are designed matters more than which algorithm is used.

▮ SIGNALThe first to feel they will be replaced are those at the frontier of research, and their response is to build the system that replaces them.

❯ Online healthcare store General Medicine raises a $120 million Series B to put prescriptions, tests and visits on one shelf

a16z leads $120 millionUS online healthcare platform General Medicine announced a $120 million Series B on October 6, led by Andreessen Horowitz (a16z) with participation from Matrix, VXI Capital, Eli Lilly, Mercy Health and BoxGroup. According to Digital Health News, including a $32 million round a year earlier the company has raised $152 million in total.

Buying healthcare like shopping onlineGeneral Medicine runs an online healthcare store where users can search, compare and buy care. It currently lists more than 2,900 products and services covering prescriptions, lab tests, telehealth and some in-person visits and procedures. The platform shows insurance and cash prices side by side, so users know the cost before ordering. The company was founded by the team behind online pharmacy PillPack, and co-founder TJ Parker is CEO. The report did not say how it charges.

AI at the front desk, clinicians in chargeUsers can shop directly, consult a clinician or describe what they need through an AI chat interface; when clinical review is required, a licensed clinician determines the appropriate care. The platform gathers an average of 11 years of health history per customer and gives clinicians a prioritized summary. The company says it has hundreds of thousands of customers across all 50 US states, and that 83% of users said the platform helped them address a healthcare need they had been putting off.

A drugmaker and a health system also investThe new money will go toward adding categories, including imaging and more outpatient procedures. The two strategic investors each have their own interest: Eli Lilly in direct-to-consumer channels, and Mercy Health in potentially connecting digital services with in-person care. The change patients notice is seeing the price before deciding, instead of getting the bill after the test.

▮ SIGNALIn healthcare, AI is first changing the process steps of finding the right service, learning the price and booking; diagnosis itself is still decided by clinicians.

❯ Pharma robotics company Multiply Labs raises a $75 million Series B to have robots take over cell therapy production

Patrick Soon-Shiong leadsSan Francisco robotics company Multiply Labs has raised a $75 million Series B led by Patrick Soon-Shiong and NantWorks, the company he founded. According to AI Insider on October 6, new investors include AstraZeneca, Lingotto and Teradyne, with existing backers such as Lux Capital and Founders Fund returning; total funding exceeds $100 million.

Handing manual steps to robot armsFounded in 2016, Multiply Labs develops autonomous robotic systems for drug manufacturing. The company says many biologic manufacturing steps are still manual, which slows production and raises contamination risk. Its robotic clusters perform these timed, repetitive steps in an enclosed environment that complies with good manufacturing practice, and work with drugmakers’ existing equipment without requiring a new facility or changes to validated processes.

It sells equipment, not contract manufacturingThe systems are bought and operated by pharmaceutical and biologics companies; Multiply Labs does not offer outsourced manufacturing. The company started with cell and gene therapies and is expanding into antibodies, viral vectors and mRNA therapies. It claims its robotic clusters cut the cost per dose by 74% and deliver up to 100 times the throughput of manual manufacturing; both figures come from the company and have not been independently verified.

From clinical to commercial productionThe new funding will be used to expand manufacturing capacity, speed up product development and grow the engineering, regulatory and commercial teams as the company moves from clinical-stage deployments to commercial production. CEO Fred Parietti said robotics is the only way to deliver these therapies to patients at scale. If robots can reliably meet regulators’ requirements for the manufacturing process, drugmakers will not need to add skilled technicians in proportion as they scale up.

▮ SIGNALWith AI speeding up drug discovery, the bottleneck has moved to manufacturing, and capital is starting to flow into the automation that gets drugs made.