We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Anthropic's growth slows on a quarterly basis, Meta releases open weights with caveats, and AI security draws M&A and nonprofit money
•
6 min read
•
2402 docs
This period: a closer look at Anthropic's Q3 numbers, Muse Spark Open benchmarked against Chinese open models, Cohere and Aisle's cyber deal, a16z's SAIF, Ramp at $60B, and evidence of froth in early-stage pricing.
Anthropic: strong quarter, but growth is slowing
Last period's brief covered the Q3 figures themselves. This period, Exponential View looked at the trend behind them. Using the leaked investor briefing, it reports $16.8bn in Q3 revenue, $1.70 earned per $1 of compute, an 85% adjusted gross margin, and an 11% adjusted operating margin, even though compute for R&D took 44% of revenue . It still calls the quarter a miss, because expectations had been set by a July leak of $65bn annualized revenue. Quarter-on-quarter growth dropped from 147% (Q1→Q2) to 45% (Q2→Q3) . Reaching $60bn for the full year would take $26.9bn in Q4, which is 60% sequential growth. If Q4 grows at Q3's rate, the year ends at $57.5bn . The same piece says the KOSPI fell 7.57% on the day and that AI-linked tech stocks were set to open sharply lower .
Benchmark partners pushed back on reading this as weakness in AI demand. They say their portfolio companies have customers ready to pay at good gross margins but cannot get enough compute. They see the bigger risk as market swings driven by exaggerated online claims, and say OpenAI and Anthropic going public would help by putting real numbers out .
Meta's open-weight release looks weaker on independent tests
Meta released Muse Spark Open, which it calls the most capable US open-weight model to date. It has 816B total / 42B active parameters, a 1M-token context window, and native multimodality . Independent results are more mixed:
- Artificial Analysis scores it 39. That is first among US open models but behind GLM-5.3 (45) and Kimi K3 (44). It ranks last among the top nine open models on agentic evaluations, even though Meta pitches it for agentic work .
- The license requires attribution and a separate agreement for model-as-a-service businesses with more than $20M in annual revenue .
- Launch pricing is $1.25/$4.25 per million input/output tokens. The NVFP4 checkpoint is about 530GB and runs on two 512GB Mac Studios .
The frontier model, Muse Wonder Full, is only available inside Muse Code. That has led investors to call its benchmark results "performative" for builders . Separately, the FT reports that Nvidia is in early talks to acquire Reflection AI or deepen its stake. Nvidia has already invested $800M, Reflection last raised at a $25B pre-money valuation, and the rumored price is around $30B .
AI cyber defense: acquisitions, nonprofits, internal reallocations
- Cohere + Aisle. Aidan Gomez announced Aisle is joining Cohere; the linked blog describes the deal as a letter of intent. He credits Aisle with 200+ high or critical CVEs in curl, FFmpeg, FreeBSD and Wireshark . He also says it found 29 new curl CVEs where Anthropic's Mythos found zero . Cohere plans private and air-gapped scanning for critical-infrastructure companies . Both claims are Gomez's, made while promoting the deal.
- SAIF. Anne Neuberger launched a nonprofit that provides AI cyber defense at no cost to small cities, school districts and rural hospitals. a16z's founders committed $25M .
- OpenAI. Greg Brockman said OpenAI put about 25% of its production engineers on security only. They used its top model to red-team OpenAI's own infrastructure until it stopped finding new issues . He describes the current period as a "defender's window," cites model-driven bank hacks in South Korea, and points to a $1B commitment to give defenders access to OpenAI's models . He thinks security spending should "vastly outstrip" software spending, though he adds he may be wrong .
Capital markets: froth and record valuations
Ramp raised $1.85B at a $60B valuation . Keith Rabois says revenue has accelerated for seven straight quarters .
Decacorns and big rounds. PitchBook counts 42 new decacorns in 2026 so far, above the 2021 record of 34 . Other deals discussed on 20VC:
- Isomorphic Labs is in early talks at $40B or more, five months after a $2.1B Series B .
- Mecka took a $60M Sequoia Series B. It sells data from sensor-wearing workers to humanoid developers, reported more than $100M in run-rate revenue in June, and expects $300M by year end. The caveat: one logistics company counted 42 humanoid proofs of concept and zero deployments .
- Jev raised $870M at $7.5B, 24 days after launch, in a round led by a16z . Jason Lemkin, who uses it in production, says it is right about 70–80% of the time on carefully prompted decisions .
Series A pricing. Disrupt panelists said traditional Series A metrics no longer apply. One called seed and venture more "dysfunctional" than they have ever seen, with the same company sometimes closing at two or three different prices on the same day . Their example: Dominion Dynamics, a hard-tech drone startup, raised a $100M Series A five months after its seed. The round was oversubscribed at $150M and cut back, without fundamental metrics to support the valuation .
Valar / Day One correction. Day One did not sue Valar. Valar filed in Delaware for a declaratory judgment after Day One refused to give up allocation under its spring-back pro rata clause . Neo says it stopped taking pro rata rights this year and cited this dispute .
Low-multiple exits. Bending Spoons bought Miro at about 2.3x ARR and Airtable at about 2.7x, 37 days apart: $1.08B of ARR for $2.64B of enterprise value. It underwrites deals to a 25% unlevered IRR, and Miro was valued at $17.5B in 2022 . Its target list has more than 1,000 companies . For founders of mid-sized SaaS companies, this sets a pricing floor based on free cash flow.
Agent infrastructure and payments
VGS, which holds payment credentials behind about a third of e-commerce, went from zero agentic tokens to a few million a week in three months. Stripe Link, Shop Pay and PayPal are racing to connect to agents, and card networks are moving toward open agentic-token standards . Wallet distribution is large: Link has 300M users, Shop Pay 250M, PayPal 400M+. But virtual-card flows can break the merchant relationship, for example when an agent buys at Walmart and the customer then tries to sign up for Walmart+ .
Perplexity open-sourced SPACE, its Apache 2.0 Firecracker-based agent sandbox. None of 108 escape attempts by models with root inside the VM reached the host .
Deep tech
- Recursive Intelligence. Founded by the AlphaChip creators, it has raised over $370M at a $4B valuation . It is building a pipeline from AI workload to fab-ready GDS . The founders say Nvidia described a 1,000x speedup Recursive achieved on a key design stage .
- Agility Robotics. The founder confirmed on stage that the company is about to go public, after more than ten years of building .
- Interpretability. Goodfire's Nick Cammarata claims most neurons in models up to trillions of parameters are legible, compared with about 2% in his earlier OpenAI work . Neel Nanda says the claim doesn't change his core views, and the paper is not yet published .
Anthropic's growth slows on a quarterly basis, Meta releases open weights with caveats, and AI security draws M&A and nonprofit money
20VC with Harry Stebbings
- Anthropic’s Haiku 5.5 is described as nearly as capable as Sonnet for many use cases at roughly one-tenth to one-twentieth the price, while Sonnet 5.5 is reported as twice as fast and 10% cheaper than the prior Sonnet; the speakers call these practical inference-cost deflation signals. A 90% price cut only preserves spending if usage grows about 10×, and Haiku 5.5 is said to be slower than its predecessor, leaving realized economics dependent on usage and latency.
- Meta’s Muse Wonder was described as just shy of Opus 5.5 on an independent index, but access was disputed: one speaker said it was internal-only with no usable API, while another noted an API had reportedly been announced but had not verified it. Despite the access caveat, the panel saw potential for a major Meta product through the Muse app, citing the model’s capabilities, autonomy, and Meta’s investment.
- Nvidia was rumored to be buying open-weight model builder Reflection for $30B, after a last round reportedly valuing it around $25B. The panel’s strategic thesis is that Nvidia benefits when more AI spending goes to open models running on its chips, but its own model efforts could become formidable competition for independent providers.
- An investment-committee segment describes Jev as raising an $870M Series A at a $7.5B valuation weeks after launch, but the panel calls the pitch a parody and then says it is not a parody, so the financing should not be treated as a cleanly verified deal report. The underlying investment thesis is a fast, cheap classifier that currently gets roughly 70–80% of carefully prompted matches right and may serve about 10% of AI use cases; the upside depends on improving accuracy against competing offerings from Cloudflare, OpenAI, Anthropic, and Amazon.
- The Isomorphic Labs discussion frames AI drug discovery as a different value-capture model from labor automation: the large upside would come from royalties on successful drugs, not merely saving research wages. The speakers see substantial potential but caution that clinical trials and the difficulty of securing meaningful royalties could constrain value capture.
- The episode cites a $60M Sequoia Series B for a robotics/physical-world-data provider and strong early demand for training data, but argues that billion-dollar market scale depends on robot deployments, which remain unproven.
- The panel reports 42 startups reaching decacorn status in 2026 so far, versus 34 in 2021 and roughly 20 per decade across the previous five decades; it says this could signal either unprecedented transformation or valuations getting ahead of fundamentals, while acknowledging market-cap growth. Separately, it argues that the SaaS rebound mainly benefited companies above $1B in revenue, while smaller vendors still face weak demand and a safe IPO may now require roughly $1–2B ARR rather than the previously assumed $500M.
TechCrunch
- Recursive Intelligence co-founders Anna Goldie and Azalia Mirhoseini previously developed AlphaChip at Google Brain with Jeff Dean; the interviewer says AlphaChip demonstrated reinforcement learning for chip design and was used internally for Google TPUs, and describes Recursive as having raised over $370 million at a $4 billion valuation.
- Recursive is building an end-to-end platform that takes an AI workload through accelerator-architecture design to fab-ready GDS. It combines LLMs with other AI for massive combinatorial optimization—a task the founders say LLMs do not handle well—using fast, deterministic methods and verifiable physical rewards such as area, power, and clock frequency; the methods are proprietary. The founders also say it can optimize flattened chips of hundreds of millions or billions of cells rather than separately partitioned blocks.
- The founders say Nvidia discussed Recursive at DAC after the startup accelerated a key chip-design stage for Nvidia by 1,000×; they also say nearly every major semiconductor company has reached out and that Recursive works with semiconductor companies, frontier labs, fabs, and memory providers.
- The founders' economic case includes a hypothetical: if a new AI architecture achieves only 5% utilization on existing hardware, a custom chip could have about 20× room to improve and use 20× less compute for equal quality; they project compute demand to grow at least 10,000× over five years. Their turnkey plan is for companies new to chip design to provide workloads and performance specifications while Recursive handles design, fab coordination, and packaging, though the founders acknowledge fabrication is a bottleneck.
TechCrunch
- Apate AI sells conversational agents that engage scammers across voice, text, email, and web, extracting financial and technical intelligence for banks, telcos, and investigators. The company reports 9.3 million autonomous conversations, more than $33 million in avoided scam losses, and a 76% reduction in losses at a major Australian bank; it says sales are harder in markets without an urgent regulatory or financial incentive to act. It says it is in production with major financial institutions and government agencies across multiple countries.
- Cedar DB announced managed CDB Cloud for its single database engine serving transactions, analytics, and AI on live data. The founder claimed more than 60,000 transactions per second alongside analytical queries under 100 milliseconds on a low-tier instance, and reported over 2,000 on-prem projects started in the prior month; however, adoption to date was described as mostly community-edition use, with cloud intended to serve users seeking a hosted option.
- Prism is building bank-integrated inheritance infrastructure to capture beneficiary relationships before wealth transfers; its pitch cited $124 trillion expected to shift to the next generation and a survey finding 70% of financial institutions unready to capture those account holders. It reported being live with a $37 billion Mississippi bank, integrated with all three of its Jack Henry cores, with 39 institutions in its active pipeline and access to 1,700 through the integration; deployments are pitched at 12 weeks, with annual and transaction fees.
- Flocker automates customer-data cleaning, pipeline maintenance, and business-entity modeling to give agents and teams a connected data foundation; it also announced plans to go open source. The company says hundreds of firms use it, including Perplexity, Wise, and AngelList, and claims millions of dollars in customer pipeline, over 70% less time spent on customer research, and recent 14% month-over-month compound revenue growth. It says customers start around $10,000 annual contract value, with some expanding to six figures; its stated differentiation is the agent context for choosing among 80-plus data vendors.
- GDA says its platform tracks narratives across 600 social platforms and looks for coordinated or inauthentic behavior before volume spikes; it lists Fortune 50 media and retail companies, a frontier AI research lab, and a major music group among its customers. A demo linked early Japanese-forum activity to a later chip-sector sell-off, but the presenter explicitly did not claim the forum drove the market move. In Q&A, GDA said brands decide how to respond, while its platform identifies organic versus inorganic posting behavior and supplies data for follow-up.
a16z
- Agentic commerce is moving from recommendations to completed tasks: one example involved an agent finding piano teachers, checking schedules, and booking a first lesson; the interview links agents’ increased utility to payment credentials and says wallets are racing to connect with agents.
- VGS says it secures credentials representing about one-third of e-commerce GMV and holds 10 billion credentials ; it reports agentic-token provisioning rose from zero to a few million per week in three months. The described card-network model lets a wallet pass a credential to a verified agent with a spending limit, so the merchant charges the consumer’s underlying card; the interview says this can preserve merchant follow-on transactions and chargeback protections. By contrast, a Stripe Link one-time virtual-card purchase at Walmart showed as a Stripe transaction and failed when the customer tried to pay for a Walmart Plus subscription. VGS frames credential brokering among parties as a new “network era” built on its vault and credential-update services.
- Adoption and liability remain unsettled: the interview attributes Amazon’s blocking of Muse partly to advertising incentives, while suggesting Walmart and Shopify merchants may have different incentives to accept agents. It also says responsibility for agent mistakes remains unresolved, even when the agent and its instruction are verified.
Fei-Fei Li
Profile
- The event introduction reported that AMD had acquired Fei-Fei Li’s World Labs “last week,” but its wording—“Neo Frontier Lab World Labs”—is unclear, so treat the deal identity as an unverified lead.
- Panelists said AI tools had cut hardware-design time in half across verification, logic and synthesis, and placement and routing; they also argued that capable generative-AI models can act as quasi-products, shortening the path from idea to practical or economic value.
- Li argued that open models support entrepreneurship, knowledge sharing, and evaluation; Hennessy said they had narrowed the gap with frontier models and might be available at about one-tenth the cost for many uses, though he noted gaps in coding and some other areas. Li also argued that universities, unable to match the resources needed to train large language models, should build on distinctive data, expertise, and problems rather than compete directly with Google’s LLMs.
- Li said AI and software needs are spreading into healthcare, energy, and manufacturing, creating opportunities beyond conventional tech employers. Speakers also described open-model guardrails as a research challenge: safeguards can be removed, and they said there was no good solution yet for the biosecurity risks raised.
Demis Hassabis
Profile
- Google held back its early language model LaMDA amid concern that a frequently inaccurate chatbot could undermine Search’s reliability reputation; the interview contrasts this with OpenAI’s greater startup latitude to take launch risks, and says low-stakes uses such as summarization and homework revealed unexpectedly strong demand despite the models’ flaws.
- The interview describes the Google Brain–DeepMind merger as overcoming geographic separation and prior internal competition, and reports that Gemini 3.0 was ahead of a rival on independent power and accuracy leaderboards by the end of 2025.
- Hassabis’s AI-for-science thesis is to train models on domain data and use them to guide search through vast combinatorial spaces, rather than enumerate every option; he says this approach is useful when a problem has a clear objective. He prioritizes medicine and disease, alongside environmental and energy applications, and cites work including AlphaFold, Isomorphic Labs, fusion, weather prediction, climate models, and materials science.
- Hassabis describes AI as dual-use and says the commercial race must not displace the priority of making AGI safe and beneficial; he calls for international cooperation on safety, including among leading labs in China.
TechCrunch
- Wabi’s CEO described a physical-AI stack combining vehicle-driving intelligence with a safety simulator, cited Volvo as a partner, and said an Uber robotaxi partnership was intended to deploy more than 25,000 taxis. She said Wabi had raised $1.4 billion without building demos, arguing that demos disconnected from scalable products can defer problems and that the safety bar rises with deployment scale.
- An autonomous-vehicle panelist warned that investor pressure and limited runway can contribute to putting incompletely validated driverless vehicles on public roads, and said their company would not do so. Another panelist recalled that some demo-focused AV companies from 2017–18 failed to reach Series B and said founders should plan the transition from demonstrations to scalable products early.
- A Shield AI speaker said the company had deployed systems in the air, at sea, and in space, and invested more in trustworthiness and reliability infrastructure than in edge-level intelligence. The panel described keeping non-deterministic capabilities outside critical feedback loops, with runtime assurance and mitigation; for Level 4 AVs, it also highlighted fault diagnosis and autonomous minimum-risk maneuvers.
- In industrial robotics, a panelist said applicable safety regulations prohibit non-deterministic algorithms from safety functions; model-driven humanoid control introduces fall and flailing hazards, prompting customer scrutiny and some companies to use wheeled designs or simplify robots around their hands.
TechCrunch
- Panelists said AI-era Series A rounds lack a reliable ARR or growth cutoff: investors increasingly look for a credible path to a very large category outcome, a persuasive narrative, a high-caliber technical team, and evidence the product works. One cited portfolio example, hard-tech autonomous-drone startup Dominion Dynamics, raised a $100 million Series A five months after its seed round; the panelist said it was oversubscribed at $150 million and cut back, despite limited fundamental metrics to support the valuation or speed of the raise.
- One investor described early-stage pricing as frothy and highly dispersed, with the same company potentially facing two or three different prices on the same day; another called the market unusually optimistic and predicted a correction in 2027. Panelists also anticipated a pronounced haves-and-have-nots market, where some companies struggle to raise while others secure very large rounds.
- Investor attention is concentrated in defense and other hard problems, including precision manufacturing, transportation, chips and GPUs, and photonic interconnects, as well as AI infrastructure such as data, inference, and chips. Panelists cited AI capex and defense-sector tailwinds, while noting defense businesses are capital-intensive and government procurement is long and lumpy. For frontier-model companies, they described heavy compute spending, negative gross margins, and rapid model depreciation; one investor argued that underwriting should compare current training spend with potential next-generation revenue and assess gross-profit scale rather than assume traditional software margins.
- Revenue diligence red flags include cumulative revenue or ARR charts and presenting a month's net-new revenue multiplied by 12 as ARR; a panelist urged conservative reporting and warned of false positives amid unusually strong VC optimism. Foundation-model labs also pose a competitive risk as they move toward applications and compute, though panelists saw room for new cloud and lab entrants with a well-chosen wedge and advised considering the likely long-term industry structure.
- One investor's historical portfolio analysis found that first-time versus repeat-founder status mattered less than team composition: teams with multiple cofounders and at least one technical cofounder were the best-performing archetype, while solo non-technical founders were the lowest-performing.
Elad Gil
Profile
- OpenAI president Greg Brockman describes a temporary “defender’s window” in cybersecurity: frontier AI can help defenders find and fix vulnerabilities before comparable offensive capabilities become broadly available; he cites high-profile South Korean bank hacks attributed to models. OpenAI says it provides capable models to defenders through Daybreak and has committed $1 billion to frontline defenders. Brockman says OpenAI paused other projects for about 25% of its production engineers to secure its systems, and argues cyber and security spending should vastly outstrip software spending, while acknowledging he may be wrong.
- Brockman identifies drug discovery as AI’s clearest candidate for major transformation, but says access to sufficient biological data remains unresolved; he points to regulatory filings—which he says can take about a year—as a workflow bottleneck AI could ease, while data gaps will require wet-lab and simulation innovation.
- Brockman reports that OpenAI published nearly 400 important open math problems that have been resolved, with mathematicians reviewing the work for insights; he presents this as evidence AI can generate new knowledge, while saying models are not yet producing wholly new theories.
- Brockman expects AI to lower startup barriers, citing people he knows who are leaving jobs to found companies because reliable AI can take on work that previously made starting a business harder; he also expects consumer and workplace assistants to converge, although compute costs will continue to shape their differences.
Greylock
- Medra is building a “physical AI scientist” for drug discovery: a robotic platform that learns scientists’ tacit lab practices, reproduces and generalizes protocols, and connects biological reasoning and assay designs to robot-executable experiments. Its approach uses modular agents for multimodal lab data and a continuous-learning loop to adjust experiment parameters and learn from outcomes.
- Medra identifies bottlenecks to closed-loop experimentation: pharma and biotech data can be extensive but unsanitized and missing usable experimental metadata, while lab instruments often lack API access and sufficient automation. It also says transferring hypotheses from dry labs to wet labs or CROs fragments context and loses process know-how, such as parameter adjustments.
- Medra reports that it built an autonomous lab in roughly 75 days, with 100+ robots running science 24/7, and has closed-beta partnerships with foundation-model companies and government research agencies.
TechCrunch
- Gamma’s founder said its first 1,000+ signups, mostly from friends and the founders’ network, were weak validation: friends praised the product but often did not use it. A public Product Hunt beta provided more useful customer feedback and won product of the day, week, and month.
- The panel recommended starting with a narrow customer niche and founder-led outreach, including events and direct conversations. AI can add leverage, but founders still need human-to-human qualitative customer discovery, and most startups should not count on strangers flocking to a viral launch.
- The panel’s early product-market-fit indicators included recurring or increasing use, referrals, customers who chose the product after considering alternatives, and passionate users. For AI consumption products, it also cautioned investors to test whether customers stay when discounts are removed and margins become viable.
- The panel framed early AI moats as unusual rather than assumed: a VC panelist cited exclusive data access or explicit regulatory leadership as possible but unlikely early advantages, and emphasized execution and rapid iteration. Customer deployments may also provide data for model training or custom workflows.
- Creator marketing should rely on authentic product users, with incentives judged partly by whether creators keep using the product if payments stop; otherwise, acquisition may be unsustainable. A panelist also said online discussion can influence discovery through AI models such as OpenAI, Gemini, and Claude.
- For B2B startups, moving from a mature sales-led motion to product-led growth can be difficult and may cannibalize larger deals. Self-service or consumption-based sales can fit customer segments too small to serve profitably with human sales, provided the product substantially eases onboarding.
TechCrunch
- Agility’s Digit V4 is deployed at customer workplaces, and the founder confirmed it is already generating revenue. Its initial commercial focus is bounded warehouse box workflows, including stacking and transferring boxes between mobile robots, conveyors and shelves, rather than general-purpose automation from day one.
- The founder said moving from proof-of-concept robots to commercial products required 3–4 years of practical trials and two years operating a robot to learn hundreds of specific customer requirements. He characterized deployment, maintenance and reliability as difficult execution challenges and a barrier to competitors.
- Agility announced Digit V5 as designed for operation beyond physical barriers. Its safety approach includes safety-rated controls and an independent supervisory computer, reliable detection of nearby people, and behaviors such as slowing down, setting down its load or sitting and cutting motor power; Agility is also involved in drafting ISO standards for dynamic-stability safety.
- The interviewer described Agility as on the verge of going public, and the founder confirmed it, calling the IPO a starting point for the next growth phase. He said the company had spent over a decade methodically building its technical, organizational, product and customer foundations.
No Priors: AI, Machine Learning, Tech, & Startups
- OpenAI co-founder Greg Brockman said nearly 400 important open math problems had been resolved and were being reviewed by mathematicians for insights ; he described the results as evidence that AI can generate new knowledge, while noting models are not yet producing wholly new theories .
- Brockman called drug discovery the clearest target for AI-driven scientific transformation, but said access to biological data remains unresolved and wet-lab and simulation advances may be needed . He also cited regulatory filings as a bottleneck AI could ease, while arguing changes to clinical-trial requirements should depend on evidence that models can predict outcomes effectively .
- Brockman described Dots as an always-on work agent that tracks tasks and integrates context across his calendar, inbox, and Slack . He expects consumer and workplace assistants to converge, with compute cost shaping product differences, and said AI is lowering barriers for people starting companies .
- Brockman said cyber capabilities available to more users are rising and cited high-profile South Korean bank hacks driven by models; he described a current window for defenders to use frontier AI to find and fix vulnerabilities . OpenAI has made a $1 billion commitment to help frontline defenders, and Brockman said it reassigned about 25% of its production engineers to security work, using models to find and fix issues . He argued cyber and security spending should vastly outstrip software spending as software gets cheaper and resilience becomes more valuable .
TechCrunch
- Field AI describes its product as a general-purpose intelligence layer for robot forms ranging from quadrupeds and humanoids to multi-ton trucks, aimed at construction, mining, energy, data centers and other difficult environments where connectivity, GPS or pre-mapped routes may be unavailable. The company says its robots are deployed across four continents at construction, oil-and-gas and delivery sites.
- Field AI’s fleet-learning loop sends field data to the cloud when robots return to charging stations and have connectivity; the resulting model is adapted for edge deployment, after which robots operate autonomously offline. The company describes this as a recurring daily learning cycle, with larger feature releases sent to robots quarterly.
- Field AI argues that physical AI cannot rely on scale alone: it says the largest physical-world datasets are still in the millions, versus hundreds of trillions of tokens for digital AI, and that physical errors have direct consequences. Its stated approach combines data-driven learning with physics-, risk- and uncertainty-aware methods so robots can detect possible hallucinations and slow down or reassess.
- Eclipse partner Aidan says the firm manages about $12 billion in assets and cites portfolio companies spanning space, construction, defense and autonomous vehicles, including True Anomaly, Bedrock Robotics, Gambit and Wave. He also flags real-world evaluation as a major deployment hurdle: many companies remain in the lab, and success on lab metrics may not carry over to unpredictable field conditions.
TechCrunch
- Panos said Amazon has more than 500 million connected Alexa endpoints and put Alexa+ usage at “a couple hundred million”; he said people talk to Alexa 2–3 times more than before.
- Amazon introduced B Frontier and B Eclipse, all-day ambient-listening devices that can surface conversation summaries, to-dos and commitments with dates, plus signals such as fatigue or stress. Amazon’s stated privacy promise is that the data stays with the user, is not seen by B or Amazon, and is not used for other purposes; sharing or adding connectors requires permission.
- Panos expects smartphones to persist rather than be replaced: tasks will continue to shift to whichever form factor suits them, while ambient devices complement existing devices and an intelligence layer connects them.
- Amazon said its Leo satellite service had more than 1,000 satellites built and roughly 396–400 in orbit, with more launches planned over the following months and service expected soon; its stated plans include broadband and direct-to-device support for Apple’s SOS services in coming years.
- For founders, Panos argued that products win by making life easier rather than requiring behavior change, and that teams should test their products through real use. He cited Microsoft’s $1.2 billion Surface RT write-down and earlier software, inventory, and retail decisions, and advised learning from failure, holding onto a valid vision, and adapting and moving quickly.
Lightspeed Venture Partners
- Aledade’s value-based-care thesis pairs independent primary-care practices with payer contracts that reward improved quality and lower total-cost trends, using technology and prevention to change practice economics. The interview says ACO value-based payments now total billions of dollars and the Medicare Shared Savings Program covers 40% of traditional Medicare, suggesting a materially larger incentive base for prevention than existed when Aledade began.
- In healthcare AI, voice-agent capabilities may become commoditized; the interviewee points to distribution, proprietary data and differentiated business models as potential advantages, and argues that patient trust may favor outreach through a familiar primary-care practice and a doctor’s voice.
- Aledade used a frontier visual-language model and then trained its own open-source model for a specific workflow: applying decision support across different EHR interfaces. The interviewee sees domain-specific, owned models as a recurring pattern when builders know the use case well.
- The interviewee considers USMLE-style performance an unreliable proxy for clinical ability because benchmark vignettes omit the fragmented, ambiguous information and repeated interactions of real care; he sees scribe conversations as useful raw material for consumer health chat that connects patient and clinician contexts.
- Mostashari argues that experienced healthcare founders, including physicians entering entrepreneurship later in their careers, are underestimated.
TechCrunch
- Bluesky says its protocol upgrade for private data is in alpha and rolling out, with access planned for everyone by year-end to meet demand for private communities; it reports thousands of apps and developers building on the protocol.
- Bluesky’s Addi/ADI is adding an AI-assisted builder for sites and social experiences, letting people without technical backgrounds create for groups without building social infrastructure; the team says builders can reach Bluesky users immediately.
- Bluesky’s exploration team is investigating credentialed human-agent interactions and a ledger of activity; the interview describes DelTown, launched 10–12 days earlier and said to be going viral, as agents communicating on the protocol.
- AI adoption is a caveat: the interviewer said Addi became Bluesky’s second-most-blocked account after JD Vance; Bluesky acknowledged user opposition to AI while arguing for AI-assisted building and saying it tries to keep unwanted AI content off the network.
- Bluesky plans to begin monetization in the coming months with payments supporting commerce and fundraising across the network. It presents its PBC and open-network approach as an alternative to lock-in: users can move their identity and social graph to other services.
TechCrunch
- Les Karpus, NVIDIA Inception’s global head of physical AI, says he prioritizes startups that understand customer problems as deeply as their technology, demonstrate product-market fit, and can integrate NVIDIA’s open libraries, models, SDKs, and tools to deliver customer ROI without heavy guidance.
- Physical-AI defensibility depends on distinctive data and model weights: even companies using the same stack may develop different capabilities from different customer bases and data. He favors combining high-quality, metadata-rich collected data with task-specific synthetic data, and sees reusable 3D simulation data as a safer long-term investment. He also advises startups seeking durable value to own the model and post-training at the core of their offering rather than rely on a major foundation model.
- Task-focused startups may reach ROI sooner, but Karpus expects general models to perform specific tasks equally well and dominate after an inflection point he places roughly one to three years out; deployed startups should build on their real-world deployment weights and brand to expand into adjacent markets.
- He estimates that much humanoid-robot ROI is three to five years away: general-purpose humanoids face factory regulatory hurdles and residential liability concerns, while wheeled designs are easier to deploy today; legs remain useful on stairs and uneven terrain.
- A key progress test is whether robots can learn unfamiliar tasks in new environments and recover from their own errors, alongside growth in commercial deployments. Off-shift automation in factories or construction could still improve productivity even without rules allowing humans and robots to work side by side.
TechCrunch
- Hello Robot’s Stretch is seeing use outside the lab: about 350 robots are in the field, including at research labs, and two are being tested for daily home use by people with severe motor impairments. The system combines autonomous navigation with user control of final manipulation, since robots can still fail in unstructured home settings and some users value being in the loop.
- Hello Robot says Stretch’s nonhumanoid design has 9 degrees of freedom versus roughly 20–40 for a humanoid, making it less complex and safer, with a 5–10× lower cost floor; the current research and enterprise price is $30,000. A home product for people with severe motor impairments is planned for a future generation after Stretch 4, with further robustness and certification work; the company identifies payment as a major obstacle, with insurance and reimbursement potentially taking years.
- The assistive-home market is the company’s intended beachhead: it describes the market as small but strongly motivated, and says deployment also requires support, including occupational therapists training caregivers. Natural-language control is demonstrated but not yet productized, while the company’s CEO argues that robotics will take longer than a ChatGPT-style transition because hardware and data requirements differ; he expects foundation-model capabilities to commoditize and sees durable advantage in solving customer problems with a complete offering and a data moat.
TechCrunch
- Benchmark announced $2 billion across two vehicles; a partner said its core founder-focused early-stage fund was being allowed to grow after staying at $500 million for 20 years, while the separate T1 growth strategy would remain concentrated—roughly four to six investments, possibly seven to eight—with a targeted 10–100x return profile and potential for 1,000x outcomes.
- Benchmark backed Manis, a consumer AI application built on OpenAI, Anthropic, and Google APIs with computer- and browser-use capabilities; its six Singapore-based cofounders took it from zero to a 100 million run rate in eight months, and Meta acquired it in December 2025. The investor said the acquisition closed and he then left the board.
- Benchmark partners said their portfolio companies are seeing paying-customer demand at good gross margins but cannot obtain enough compute to serve it; they described market volatility and unreliable or exaggerated online claims—not weak underlying AI demand—as a key risk, and said public reporting from OpenAI and Anthropic would help.