• AI data-center financing: The discussion said CoreWeave’s unsecured bonds’ implied yield rose from 9.6% to about 15%, while the company continued paying its 9.6% coupon; the repricing could make future borrowing more expensive. All-In Podcast Panelists estimated another ~$35B of debt was needed to finance signed contracts, said three customers represented 75% of revenue, and identified the reported 2.5-year payback period as a key risk metric if it lengthens. All-In Podcast
• Compute and power constraints: Panelists said compute demand remained strong, including premium rental prices for older H100 and A100 GPUs, while construction and local permitting constraints limited new powered capacity. All-In Podcast One argued that land, power, and data-center shells could become a moat—and that startups could be unable to launch if hyperscalers lack capacity. All-In Podcast
• Startup economics and product edges: Calacanis described early-stage AI startups as raising capital, winning customers, and reaching unusually fast revenue ramps, citing doubling revenue each quarter as a benchmark. All-In Podcast Panelists argued that coding assistants make bespoke software more viable and may expand software creation even as prices fall; they also argued coder demand can rise despite assistants writing most code, and identified proprietary vertical data and user feedback as potential moats. All-In Podcast A specialist-model example: Coinbase fine-tuned an open-weight 9B-parameter model for fraud detection, reporting about a 30% reduction in fraudulent transactions in testing while being faster and cheaper than Opus 4.5; the panel framed workload-specific small models as an emerging enterprise trend, especially under compute constraints. All-In Podcast
• Anthropic growth and financing: Panelists cited investor-day figures of $16.8B Q3 revenue, 900% net revenue retention, positive free cash flow, and adjusted gross and operating margins of 85% and 11%; Anthropic had not disclosed ARR, and secondhand estimates ranged from $70B to over $100B, with the adjustment methodology also undisclosed. All-In Podcast Chamath argued Anthropic should raise $200B–$250B, potentially at a lower valuation, to secure compute runway through a 3–5-year period; Horowitz agreed financing risk could be the main constraint but said he could not judge the precise amount needed. All-In Podcast

• Flock Intelligence is pitching edge-based wildlife AI that identifies species, age, sex and behavior, then selects context-specific bioacoustic signals; the company puts its initial infrastructure-focused wildlife-conflict market at about $15B. TechCrunch It reports deployments across three continents, including Denver International, Parks Canada/USDA, Portuguese highway operators and Swedish armed forces, and says it has validated its train product with Alstom for two years and is preparing to commercialize together. TechCrunch The founder estimates train-animal collisions cost at least $200,000 and says avoiding one or two collisions could repay the system. TechCrunch
• Apate AI deploys conversational agents across channels to engage scammers and extract intelligence—such as mule-account details, phone numbers and malicious URLs—for customer defenses. TechCrunch The company reports more than 9 million conversations, $33M+ in avoided scam losses, and customer outcomes including a 76% loss reduction at a major Australian bank, $183,000 saved by disrupting a live campaign, and mule accounts identified 61 days earlier. TechCrunch Banks are a core buyer, with modular SaaS pricing; telco pricing is consumption- or minutes-based. TechCrunch
• Intero Bio Systems makes miniature human-intestine organoids with multiple intestinal cell types and organ-level functions, selling them for compound efficacy and toxicity testing; it also positions the tissue as a validation layer for AI-designed drugs. TechCrunch It reports nearly $500,000 revenue from eight customers, strategic investment from Merck, more than $4M in non-dilutive funding and a 10x manufacturing pipeline; it is raising a $6M seed round and says it received a term sheet that week. TechCrunch The founder says preclinical screening has no regulatory barrier to market entry, while the company is discussing with the FDA how organoid data might support IND applications. TechCrunch
• Industrial Next sells AI and a sensing/computing stack for dexterous factory robotics, primarily through recurring software licenses, with optional work-cell hardware; it says its robots have assembled just over 7 million smartphones. TechCrunch One customer example was a deployed machine moved to pre-production of an unreleased phone and producing it after 60-second commissioning; the company also announced an evaluation partnership with BMW Group Technology Office. TechCrunch Its demo used about two hours of real-world data after simulation training, but the company says task data capture to deployment can still take several days; it gave target ROI timeframes of 1–1.5 years in North America and about two years in Asia. TechCrunch
• Satellite is building software to analyze data and orchestrate workloads across compatible spacecraft, while leaving command authority with operators; it says its current market is B2B and B2G. TechCrunch It reports being live on three satellites, $500,000 invoiced in commercial and space-agency revenue, and an $8M seed round closed earlier in the year; its first test reportedly reduced telemetry by 97%. TechCrunch It quotes $4,200 per satellite per month against about $50,000 in annual savings, and targets 100 satellites before its Series A and 20% of non-Starlink satellites by decade’s end. TechCrunch Its interoperability strategy includes participation in CCSDS standardization and use of NASA-developed protocols, while the founder identifies execution speed against better-funded competitors as a key risk. TechCrunch

• Wahlberg described Hallow as one of his investments and said he chose to meet its Notre Dame graduate founder while the app was prelaunch and had no following; the founder’s faith-centered mission appealed to him more than another app builder’s track record of building and selling apps. TechCrunch Wahlberg said Hallow became the world’s No. 1 religious app and Apple’s No. 1 free app during Lent each year. TechCrunch
• Wahlberg says celebrity-driven deal flow often includes people seeking to attach themselves to his platform; he prefers seeking out capable, hands-on operators. Carey describes his role as turning Wahlberg’s intuition into investment structures that benefit all parties. TechCrunch
• Wahlberg’s Municipal-linked first big-box gym was due to open in December, combining fitness with recovery, retail, supplements and epigenetic testing; he described an aggressive growth strategy and a $99 offer including three sets of blood work and a one-hour doctor consultation. TechCrunch Carey identified AI and longevity applications as a major technology tailwind, with health-data collection and tools as a possible fit for the Municipal/gym portfolio; Wahlberg said such tools could clarify health deficiencies, while lasting results still require effort. TechCrunch
• For a Pete Rose film, Wahlberg said AI could enhance the stadium setting, but rejected simply using ChatGPT to generate a movie without real actors. TechCrunch

• Amjad reported that enterprise contracts rose from 2% of Replit’s revenue last year to 15% this year, B2B exceeded 50% of revenue, and net revenue retention was over 400%. TechCrunch He also said gross margins were positive and increasing as enterprise adoption grew, with no immediate plans to raise. TechCrunch
• Replit’s enterprise pitch is an integrated stack for building, deploying, and scaling software, including production access, databases, and on-prem deployment; Amjad said enterprise buyers prioritize security, safety, and deployment over price. TechCrunch
• Amjad contrasted a consultancy’s $10 million, 18-month estimate for an accounting-software need with a few-hundred-thousand-dollar build he said could serve a Japanese bank for 10–20 years, arguing that deterministic software will remain useful even as AI coding expands. TechCrunch He cautioned that AI agents can be harder to control and introduce subtle bugs. TechCrunch
• Amjad said prompt injection remains a major problem and Replit routes requests between frontier and smaller, roughly 10-billion-parameter models, arguing a single model for every task is not feasible on cost and security grounds. TechCrunch He also described a past incident in which an agent deleted a database, with snapshots enabling one-click restoration. TechCrunch

• Meta’s Wonder API model is priced at $4 per million input tokens and $20 per million output tokens; the stronger Wonder Full is available only through Muse Code subscription tiers of about $50 or $200 per month. Lightspeed Venture Partners The presenters framed active memory and sustained focus for long-horizon agents as its key strengths, while noting weaker AI web development than Opus 5.5 and uncertainty around competing “frontier” claims. Lightspeed Venture Partners
• Google’s Gemini model became available across its API, Antigravity, Gemini Enterprise, Gemini app, and AI Mode in Search after a period limited to vetted cybersecurity partners. Lightspeed Venture Partners The discussion highlighted legal, business, and finance workflows, vision-plus-code, and visual planning and simulation, while emphasizing that the model is not itself a robotics model. Lightspeed Venture Partners
• Apple’s $299 Home View smart display recognizes up to six people and shows personalized calendars and messages when it recognizes a person and their unlocked iPhone is nearby; it is designed for iPhone-using homes. Lightspeed Venture Partners The presenters saw privacy and ecosystem integration as potential advantages for AI in the home, but said adoption is uncertain and that Siri is not yet as capable as agentic tools. Lightspeed Venture Partners
• SpaceX is pursuing a standalone US mobile-carrier business following FCC approval of a 15,000-satellite direct-to-device constellation and an $8 billion purchase of low-frequency spectrum. Lightspeed Venture Partners The network still requires towers, compatible phones, and substantial buildout, making it a longer-term challenge rather than an immediate threat to incumbents. Lightspeed Venture Partners
• Zarya raised $5.2 million in seed funding, reported as led by Lightspeed, for real-time servicing and collateral infrastructure for structured finance and digital asset-backed lending. Lightspeed Venture Partners It has applied to charter Zarya National Trust Bank and says its platform is live with Maple, Galaxy, Kraken, and Figure. Lightspeed Venture Partners
• Lightspeed was identified as an investor in Zenzap, which launched AI agents in its mobile-first work chat; the episode reports thousands of daily business users across the US, UK, and Australia, and says employees can create an agent in under 90 seconds without code. Lightspeed Venture Partners

• Greer says its telemetry-stream product can reduce observability costs by 90% and is adding AI troubleshooting to improve reliability. In customer calls, buyers reportedly focused on the Datadog-bill savings rather than the AI reliability pitch, suggesting the cost-saving value proposition is clearer to buyers today. a16z
• Greer’s operations approach looks for new patterns in logs, then has AI inspect the related code and test possible explanations, rather than starting with an alert and working backward to root cause. The speaker says a couple of ignored signals in internal or customer environments preceded production incidents—an early, anecdotal signal rather than proof of broad effectiveness. a16z
• The interviewee reports building production data-lake auto-promotion functionality alone in about 10 days, versus an estimated six months previously. He distinguishes this coding progress from operations: query optimization can work when grounded in benchmarks, but memory sizing still requires production experiments, and AI running a distributed cluster is “not yet” ready. a16z

• Field AI was reportedly eyeing a $700 million raise at a $10 billion valuation, while a German robotics startup had recently reached unicorn status. TechCrunch
• The early-stage firm’s fund-one thesis spans robotics, materials science, and life sciences; its definition of physical AI extends beyond robots to AI-enabled physical outputs such as molecules, materials, and medicines. TechCrunch
• The investor cited industrial robotics investments addressing labor gaps at small and midsize factories, and named Andromeda Surgical as a portfolio company while discussing surgical co-pilots amid surgeon burnout and waits of months for complex cases. TechCrunch
• She described robotics as in a “Lego era,” with foundational software and building blocks being rethought; simulation is becoming more critical, while open-source safety and standards remain priorities. TechCrunch
• In her assessment, China has historically had more robotics-related research and publications and long-standing manufacturing expertise; the U.S. has a strong talent pipeline, and she pointed to U.S. companies building actuators, robotic hands, materials, and robots while stressing the need for supply chains within the ecosystem. TechCrunch
• She says her prior NVIDIA work and network of researchers and developers speed up diligence and customer or pilot introductions, including before an idea becomes a company; she positions help with talent, product positioning, and partnerships as VC value beyond capital. TechCrunch

• Latch says it began in 2020 building biotech and pharma data infrastructure for molecular and multiomics data; as coding models showed promise, it saw agents potentially changing how scientists work with data and made benchmark development its sole focus.Greylock
• Latch’s benchmarks test agents on biological data-analysis tasks using assay-specific data snapshots and calibrated experimental context, starting with Spatial Bench; it has expanded toward longer tasks simulating paper results, single-cell epigenetics and whole-genome sequencing, with roughly 200 drug-discovery benchmarks on its roadmap across modalities, development stages and indications.Greylock
• Latch says Anthropic and other frontier labs have picked up its work, which has appeared in recent model cards, and reports many pharma deployments.Greylock Its biocurity work includes benchmarking models’ refusal of legitimate research and measuring pathogenic-function capabilities.Greylock

• OpenAI introduced Dots, a personal AI assistant, at Dev Day a few weeks before the interview; its product lead described it as combining an agent that can take actions with an always-on assistant that considers how to help, maintains context and can do future work. TechCrunch
• OpenAI positions Dots around capability and trust, saying it uses GPT-6 Astra and has its own computer with the ability to connect to the user’s computer; a reported drone-delivery startup example was Dot detecting and analyzing an outage and surfacing useful resolution information before the engineer noticed it. TechCrunch A cadence-based proactive-research feature considers conversations and connected information, suggests actions, and asks for approval before acting; users can authorize more proactive behavior over time. TechCrunch
• Dots is currently available to Pro, business, business premium and enterprise users, with OpenAI saying it wants eventually to bring the technology to its ChatGPT user base of well over one billion. TechCrunch The product lead said connecting Dots to Slack with its own identity shifted much of his usage there, where he used it more than the first-party app, and said OpenAI sees this pattern among other users. TechCrunch
• OpenAI describes trust controls including ChatGPT privacy controls, connector permissions, an activity log, safety layers and conservative defaults that ask for permission frequently. TechCrunch More granular file-level access remains a build-out need; the product lead also said the new product’s context management is still being developed and should not be expected to be perfect immediately. TechCrunch
• OpenAI’s product lead said the company was in discussions with banks and payment processors about how agents should interact with the world, describing this as an area for collaboration across the ecosystem. TechCrunch

• A founder said he started an AI business-autopilot startup alone with AI and, in under a year, was approaching a 25m run rate; this is a founder-reported figure. TechCrunch
• AI has made building easier, but one founder said he spent six to nine months iterating before seeing early signs of product-market fit. An investor said founders’ ability to operate a fast-shipping “software factory” is now a key benchmark, even as similar products can appear within weeks. TechCrunch
• For early go-to-market, prioritize learning which customer problems and use cases are repeatable before building a top-down sales motion; three customers acquired through a similar motion for a similar problem were presented as a stronger scaling signal than ten customers acquired in ten different ways. TechCrunch
• An investor distinguished pre-market technical advantages for database or new-lab businesses from application-layer defensibility, which he said is usually built after launch through distribution, brand, virality, and ecosystem trust; he described product differentiation itself as fleeting. TechCrunch
• One investor said training, model, and hardware ventures have returned venture funding to capital-needs sizing, estimating a new AI lab may require roughly $250m–$500m because of pretraining costs; seed and later-stage labels no longer reliably signal round size, which can range from $5m to $500m. TechCrunch
• Underwrite inference resale separately: negative-margin inference may be used as an adoption subsidy, but the investor cautioned against treating it as ARR and recommended assessing lead-to-contract conversion and unit economics. He also said top-decile SaaS growth remains fundable, particularly for companies not reselling inference. TechCrunch

• AI agents are making SaaS prospects uneven: products reliant on a user interface or automatable project-management workflows are exposed, while data-rich systems of record that can protect their data and layer agents over it may be better positioned. The discussion’s view was that SaaS economics are changing, not that SaaS is simply dead. TechCrunch
• AI-native startups’ reported gross margins varied from -50% to 70% even among companies described as growing rapidly; model, deployment, data and pricing costs all matter, and companies account inconsistently for forward-deployed engineers’ implementation work in cost of goods sold. Investors should scrutinize that accounting; the panel argued lower margins could still make sense if companies capture larger markets and generate much greater revenue and profits. TechCrunch
• AI automation can reduce the number of people using a product while increasing the value delivered, weakening seat-based pricing. Companies are exploring usage, value and outcome pricing, but outcome pricing can leave vendors bearing high delivery costs; investors should examine the cost of producing a successful outcome and the unit economics behind the chosen value metric. TechCrunch
• An investor cited portfolio company Elise AI, which automates operational workflows for housing owners and operators and was said to serve one in six residential units in the U.S. and Canada. Its potential moat includes digitizing decades of previously unscanned records and learning from new data generated across its customer base. TechCrunch
• The discussion identified context, memory, proprietary data and model routing—not just access to a foundation model—as possible sources of application differentiation; falling token prices and competition among models could later improve application margins. But enterprise AI bills can be substantial: one Fortune 500 IT leader said a predicted $40 million annual Anthropic bill reached that amount, prompting closer attention to token economics and lower-cost open models. TechCrunch
• Building software has become easier, but panelists distinguished coding from securely operating complex, data-heavy software in production; they also noted that AI customers may resist long-term contracts, with some deals potentially shifting to monthly terms. Both factors affect build-versus-buy competition and SaaS sales planning. TechCrunch

• Umi presenter Manos said he worked at Google Cloud and helped orchestrate the team that became Gemini TechCrunch. He pitched Faye, an AI-engineering agent that automates evaluation setup, failure analysis, data curation, training iterations, deployment and production monitoring TechCrunch. Umi’s stack pairs an open-source base, which Manos said had over 9,000 GitHub stars, with an enterprise platform for evaluations, data workflows, governance and lineage TechCrunch.
• Manos argues companies should use frontier models for generic tasks and coding agents, but consider owning specialized models for differentiated or high-volume use cases, citing potential gains in differentiation, cost and control TechCrunch. As anecdotal demand evidence, he cited two top-10 global banks and a smaller European bank pursuing open models or internally owned AI; the banks were not named TechCrunch.
• The enterprise platform was text-to-text at the time of the presentation, although Umi’s open-source stack supported text and images; Manos said Umi does not automate the difficult work of bringing together and importing data from disparate company systems TechCrunch. In Q&A, the moderator cited Coinbase’s reported reinforcement-learning result on customer data—said to outperform Opus 4.5, with a claimed “$100 in training data”—as a challenge to the need for Umi; Manos responded that evaluation, curation and training remain effort-intensive for companies TechCrunch.

The hosts describe “sniffers” as less hands-on investors who spot people, ideas, and emerging categories before they are obvious to others; their enthusiasm and early conviction can create what the hosts call a “conviction arbitrage.” My First Million As examples, they credit a friend, Val, with prompting Bitcoin purchases in 2014 and with predicting early that stablecoins, privacy coins, and prediction markets would become major, while not knowing which project would win. My First Million

• Molecule One says it moved from software-only chemistry to a physical lab for iterative experiments, then shifted from patching workflows to making discoveries. It reports that a discovery for a large pharmaceutical company took 27 weeks last year versus 3 weeks last month, and separately characterizes its progress as roughly a 2× speedup every six months. TechCrunch
• A startup pitching gene delivery for kidney and pancreatic diseases said its demo showed disease reversal in mice, not humans. Its approach combines AI with lab screening to pursue potentially one-time curative drugs; the founder said progress still requires better data, wet-lab work, and specialist models, rather than relying on frontier models alone. TechCrunch
• Glide’s road-to-rail business serves commercial and defense customers. Its founder said Battlefield brought investor and partner inbound and more defense exposure amid geopolitical tensions; he reported close to $7 million raised and a few million more secured, followed by team and product building. TechCrunch

Interero Biosystems won TechCrunch’s 2026 Startup Battlefield, a global early-stage startup competition; Apate was runner-up TechCrunch. The champion receives $100,000 in equity-free prize money TechCrunch.

• Hugging Face said it had signed a letter of intent to join Nvidia and that the process was ongoing. It cited Nvidia’s major open-source contributions across language models, robotics and science, and said it would continue prioritizing community needs and compute agnosticism. TechCrunch
• AI benchmark results can be hard to reproduce and compare: access to model weights helps enable verification, while compute scarcity leaves open researchers less resourced than frontier labs. Eval Eval, an 800+ person research community, developed evaluation cards as a documentation standard; agent evaluations remain costly, multilayered and less robust than model evaluations. TechCrunch
• For AI startups, evaluation should reflect the specific deployment workflow and include safety requirements, while weighing frontier performance against cost and data-protection needs. A tailored harness can reach or exceed frontier performance on narrow cyber-defense tasks, but the interviewee cautioned against treating any agent as universally reliable. TechCrunch

• Battlefield judges said the eight-minute pitches gave them limited visibility into commercial activity, metrics, and team details, so they relied on the high-level vision and founder credibility. They value a concise explanation of what the product does, for whom, why it is better, and how it reaches the market—alongside the ability to move between business details and vision. TechCrunch They also said direct answers and candidly acknowledging unknowns build credibility, while evasive or incomplete answers can weaken it. TechCrunch
• The judges framed founder-market fit as central: for conceptual, founder-led bets, the founder and vision may need to come first; for other companies, lead with the business and then establish the founder’s ability to execute. TechCrunch For category-creating startups, they said founders need to show that the category can become large enough to justify the cost of creating both the category and the company. TechCrunch They view TAM as a way to describe the underlying problem—not just current spending on the exact product—and advised showing the starting SAM and path to TAM in a way that connects to the company and its founders. TechCrunch
• On Series A team requirements, one judge said investors expect some product built and see the ability to build it as essential, while go-to-market gaps can be filled later. The panel also expects founders to clarify roles and decision-making as the company grows; one judge described scaling from three people to 20. TechCrunch
• Responding to a founder building a consumer vertical agent, a judge said there are many business models but cautioned against calling one-off transactions ARR; another noted that a “run rate” may not demonstrate repeatability. TechCrunch
• One investor described a fund thesis focused on foundational industries—including manufacturing, supply chain, energy, and critical infrastructure—and called reindustrialization a 50-year trajectory; the investor also said AI is making research used to validate investment theses easier. TechCrunch
• A judge cited a past investment in an AI public-safety platform founded by a 20-year-old with personal experience of a school shooting, saying Axon bought the company last year for about $1 billion. TechCrunch

Celeno launched Reservoir, a consumer immune-cell preservation service using blood draws followed by cell isolation and cryopreservation. The service is preservation-only, not a treatment; members own their cells and decide what to do with them later in consultation with a physician. TechCrunch

The product costs about $5,000. Celeno says at-home blood draws are available in 4,000 U.S. ZIP codes, with access through partner clinics in most major cities. TechCrunch

Celeno’s CEO said Reservoir is in the FDA preservation category, which she distinguished from therapies that require approval; this is not a claim of therapeutic approval. Potential future areas she cited include autoimmunity, cancer, and cell and tissue therapies, while describing AI’s intersection with biology and medicine as part of the uncertain future rationale for preserving cells. TechCrunch

• Clay reports its largest internal AI productivity gains in software engineering, with more code and features being produced; a question-answering data-science service over company and customer data became popular internally, and the team may take it to market. TechCrunch
• Clay introduced an inbound AI SDR that can demo a product after an initial demo and improve based on what prospects understand and on new feature releases; the speaker says it could handle the full task. TechCrunch
• Gamma shifted customer support from a potentially large team to a small team of “architects” using Finn and internal tools to determine how best to engage users, enabling service beyond what the company could previously provide. TechCrunch
• Panelists favor distributed AI experimentation: let employees closest to a problem choose tools and provide access, then support projects that gain internal traction rather than relying on top-down automation picks. TechCrunch
• Startups should build product-specific eval suites and rerun them as new models arrive—one panelist describes a roughly two-week model cadence—while keeping a consistent eval set and testing both critical behaviors and end-to-end workflows. TechCrunch
• For early use cases, panelists recommend starting with the best available model and optimizing cost or building a router only after establishing that the product works; they note token costs can rise sharply with growth and that deterministic code can replace model calls where appropriate. TechCrunch
• Gamma’s Grant argues that multi-year roadmaps are impractical and favors roughly one-month planning with close customer feedback; Clay’s speaker recommends preparing for continued exponential progress while ensuring the product works if it does not, and Anthropic’s speaker suggests an ambitious goal pursued through experiments every few weeks. TechCrunch

• Clay treated LLMs as a real-time data source for its data-aggregation product, then developed Claygent for agentic workflows after recognizing the potential of the React paper—about three months later, according to its founder. TechCrunch
• Clay defines go-to-market engineering as applying engineering and data systems to find company-specific growth opportunities. Its founder said the company initially targeted technical users managing complex, spreadsheet-heavy information work rather than simplified tools for sales reps, describing that market as uncrowded. TechCrunch
• Clay sees inbound SDR agents as viable when the task has teachable success criteria, such as demonstrating a product, but says outbound SDR agents are not ready because targeting, timing, messaging, and product fit make outcomes difficult to train. Email generation, account research, and TAM segmentation are current agent strengths, though the founder said people still need to check the work. TechCrunch
• Clay’s “self-learning revenue engine” thesis is to predict actions, observe results, and update its system; its founder identified learning across many customers as a potential moat. He also said Clay’s team grew from 300 to 700 “this year” as productivity increased, while describing AI as augmenting rather than replacing staff and saying humans remain needed in sales. TechCrunch