We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
1. Funding & Deals
a16z has turned the physical-AI thesis into a dedicated $1.1 billion vehicle. The Machine Age Fund will back founders rebuilding chips, memory, networking, systems software, power, and the machines that bring AI into the physical world. Its partners’ stated thesis is that model progress is outrunning the memory, interconnect, power, and cooling beneath it, pushing strong teams from pure software into complex hardware.
The commitment is also organizational: a16z says hardware startups now represent more than 20% of its deal flow, names recent investments including Unconventional AI, Nexthop, Volta, Atoms, and Mind Robotics, and has assembled a team that includes former Intel Data Center Group CTO Guido Appenzeller and data-center veterans Raghu Raghuram and Martin Casado. The managers describe the target founder as a systems builder who can design the chip or system while planning manufacturing, suppliers, and the surrounding ecosystem; they also say first rounds can reach hundreds of millions before a product exists.
2. Emerging Teams
Maiboli is a concrete early signal for cross-language developer tooling. Its builder created a free, MIT-licensed Mac and Windows app that accepts speech in 55-plus languages or mixed-language streams and inserts concise English at the cursor, with an optional rewrite layer for organizing nonlinear speech. The builder reports five weeks of daily use by 20 teammates, more than 3,000 dictations on Gemini 3.5 Flash, and a total bill below ₹2,500—about one cent per dictation. The founder’s background is also notable for the product-led wedge: “an accountant who moved into IT,” building around a personal workflow problem rather than a conventional developer-tool pedigree. This is usage evidence, not yet paid-market validation, but it is a useful example of multilingual friction becoming a bottom-up software product.
Sentrint pairs a narrow AI-security workflow with a noisy but promising early funnel. The solo founder reports that three weeks after launch the product had 1,200 visitors, 17 signups, and six paying users; more than 60% of traffic came from Reddit, producing a reported 35% signup-to-paid rate but only a 1.4% visit-to-signup rate. The technical wedge is a three-layer pipeline: multiple security tools produce serialized findings, Claude sees the findings rather than the full codebase to reduce false positives, and a final Claude step creates prompts tailored to the user’s chosen LLM platform; scans run in ephemeral instances. The conversion figures should be treated as directional: a commenter notes that six of 17 is too small to establish pricing fit, while Reddit-heavy traffic can inflate paid conversion and crawlers can distort the visitor denominator.
Foundera is a pedigree-led watchlist item rather than a traction case. Founder Cem describes himself as a fifth-time founder who exited a previous company, has built startups since 2011, and has managed accelerator programs including Startupbootcamp. Foundera’s agents are intended to validate problems, research markets and competitors, challenge assumptions, define an MVP, connect founders, track milestones, and improve investor visibility. The product is still being built and is seeking early users, so the current signal is founder experience and category ambition—not demonstrated adoption.
3. AI & Tech Breakthroughs
Skild AI’s S1 is a potentially important robotics task-acquisition milestone, but the evidence is still partner-reported. A Lightwork segment says the system can receive one video of a person performing a task and execute it without task-specific programming, training, or teleoperation. The claimed change is economic as much as technical: conventional task acquisition required 50–100 hours of puppeteering data and weeks of engineering, while S1 is described as ingesting a demonstration in roughly 11 minutes, adapting to substitute tools, and retrying after failures.
The same segment supplies the right diligence counterweight. Robots that beat sprinting records still struggle with opening jars, folding socks, stacking cardboard, force modulation, bimanual coordination, peg-in-hole insertion, and long-horizon recovery. For robotics investors, recovery from an unfamiliar physical state is a more informative generalization test than a single athletic benchmark.
LlamaParse is moving enterprise spreadsheet extraction toward structure-aware agents. Its native mode treats spreadsheets as irregular, linked objects with arbitrary rows and columns and cross-sheet dependencies rather than as flat pages; the product uses a tuned model-plus-harness for schema-guided extraction, including dense sheets such as balance sheets. LlamaIndex’s accompanying explanation identifies the price/performance problem: the agent must ingest a large, complex interface and emit a large volume of accurate structured output. The investable layer is therefore not simply better OCR, but the harness, schema, and cost controls needed to turn messy enterprise data into reliable machine-readable state.
Evaluation infrastructure is beginning to optimize for statistical stopping. The optstop announcement targets frontier evaluations that can consume hundreds of millions of tokens, stopping trials once estimates are sufficiently precise while continuing runs where uncertainty remains. It is a small but relevant shift from treating every benchmark trial equally toward allocating evaluation budget where it changes the conclusion.
4. Market Signals
AI infrastructure demand is becoming a power, memory, and supply-chain market—not just a model market. An a16z panel cites roughly $700 billion of collective hyperscaler capex this year and says it could reach $1 trillion next year; it describes supply across key components as effectively booked through 2028, some GPUs being resold at four times their purchase price, and simultaneous shortages of power, cooling, memory, and GPUs. The panel’s illustrative economics are why specialized infrastructure is attracting capital: if a frontier model costs $3–5 billion to train and must generate about $10 billion in inference revenue, a 20% efficiency improvement could represent $2 billion—enough, in its example, to justify a custom ASIC. Execution remains materially harder than software scaling: permits, grid access, transformers, turbines, and political constraints are already pushing some GPU-seeking companies outside the United States.
Model access is becoming a strategic dependency. OpenAI says it is ending its partnership with Cursor following Cursor’s acquisition by SpaceX, with Cursor’s direct access to OpenAI models proposed to end on November 12. OpenAI says users can continue using their own API keys and its IDE extensions for Cursor. Harrison Chase’s interpretation is that model labs will build strong model-specific harnesses while blocking access from competing labs, leaving cross-model harnesses to independent providers. Jerry Liu makes the application-layer version of the same argument: companies outside frontier labs will want a mixture of proprietary and open-weight models to optimize performance and margin without being beholden to one provider. LangChain’s early support for the new MCP specification is a concrete ecosystem response. The opportunity is a neutral routing and execution layer, though its durability depends on retaining access across increasingly competitive model platforms.
Agent adoption is entering systems of record while stressing their economics and control boundaries. Linear says agents are installed in 95% of paid workspaces and that the share of work they create rose from 3% a year ago to 50%; issues with a pull request attached grew sevenfold since the start of 2026. The same analysis cautions that “share of work” is not “50% of issues,” weighting is unspecified, installation is not engagement, and leading AI companies are overrepresented in the cohort.
A separate B2B + AI operator reports agents writing roughly 40GB—about 21 million records—into Salesforce in 30 days, 99% through the API, triggering storage overages after the business had barely used the UI. ServiceTitan then cut off Podium after nine years and roughly 1,000 shared customers when Podium’s agent expanded from lead handoff into customer conversations, scheduling, job tracking, and holding the customer record. The pattern is strategic: an agent can make an incumbent system much more useful while also making it easier to replace, and pricing built for human-scale interaction can become punitive when agents write continuously. Reliability is not solved either—the same operator says a renewal agent still occasionally invented numbers after four explicit instructions, requiring human-controlled follow-up.
5. Worth Your Time
- Watch — Robots Outrun Usain Bolt, Skild’s ChatGPT Moment & Google Bids $10M for Spirit Data. The useful combination is the S1 one-video task-acquisition claim and the sharper benchmark discussion showing why mundane manipulation and recovery matter more than sprinting.
Watch — Why Top Founders Are Racing Into AI Infrastructure. The capex, component-shortage, and custom-ASIC sections provide the clearest current framing of why infrastructure is becoming a first-principles company-building opportunity.
Read — The Agents #013. Concrete cases of agent-written records, platform conflict, workflow speed, and unreliable unsupervised output.
Read — LlamaParse’s spreadsheet extraction announcement. A concise product-level example of structure-aware document infrastructure replacing flat extraction.
Thread — OpenAI’s decision on Cursor. A timely case study in how model-lab strategy can reprice the risk of building an application or harness on someone else’s models.
- AI build economics and the product bottleneck: Andrew Ng says the cost of building with AI has “plummeted,” shifting the main challenge toward deciding what to build; he advises founders to learn AI, build quickly, talk to customers, and apply judgment when iterating. He cautions that meaningful companies still require technical depth and/or deep customer insight and integration, with focused execution after broad exploration.
- AI-native talent and operating models: Ng says software engineering is the profession most affected so far, but job openings are up and good engineers he knows are busier; developers are becoming full-stack, while marketing and recruiting roles are expanding toward broader end-to-end ownership as workers combine AI and domain skills. His teams illustrate this model: marketers build web-crawling research tools, finance automates document checks and alerts, and recruiting embeds engineers to build specialized systems.
- Local and open-weight models are becoming a viable privacy-oriented stack: Ng says recent open-weight models are approaching frontier capability and are small enough to run locally; banks can use virtual private cloud or on-premises deployments, while highly sensitive workloads can remain off the cloud. He also says models change rapidly—roughly every other week—so users should avoid becoming locked into one model.
- Learn Vector targets personalized AI learning: Andrew Ng, identified as a co-founder of Google Brain and Coursera, is leading a new organization called Learn Vector focused on more customized one-to-one learning experiences. His product rationale is that AI can raise homework scores while worsening long-term retention when students offload the work to AI. The interviewer references a $100 million Coursera investment, but the exchange does not specify financing stage or lead-investor role.
- Regulatory and competitive-market thesis: Ng argues that fear-based messaging from some leading AI companies is intended to produce regulation favoring incumbents and disadvantaging cheaper open-weight or open-source models; he says this has worsened public perception and slowed U.S. AI adoption.
- AGI-hype caution: Using a broad definition—AI capable of any intellectual task a human can perform—Ng estimates AGI remains decades away because current systems cannot reliably handle examples such as writing a PhD thesis or learning to drive in a novel environment with limited practice; he notes that varying definitions and incentives can make AGI claims appear much nearer.
- Machine Age infrastructure thesis: The Machine Age fund is targeting the computer-science stack on which AI runs: chips and complete systems, memory, networking/interconnect, storage, power chips, electricity, and software for managing infrastructure fleets; it is also interested in platforms that move AI to edge or embodied devices, while excluding heavy regulated or highly verticalized industries.
- Demand and market signal: The managers describe AI infrastructure demand as structurally outpacing supply: hyperscaler capex is cited at about $700 billion this year and supposedly $1 trillion next year, supply across the board is described as booked through 2028, a leading memory vendor says current demand alone would take three years of capacity to fulfill, and some GPUs are reportedly being resold at four times their purchase price. Demand is expanding from chat and coding to knowledge-worker tools, back-office agents, computer use, and eventually embodied AI, with each progression potentially adding an order of magnitude of consumption.
- Technical opportunity: Investors see a need to redesign inference infrastructure from first principles across compute, memory hierarchy, interconnect, power, and cooling. Their illustrative economics suggest model-specific ASICs may be viable: a frontier model may cost $3–5 billion to train, require roughly $10 billion in inference revenue to pay back, and make a 20% efficiency gain worth about $2 billion.
- Teams, capital, and execution risk: The target founder is a systems builder who can architect the chip or system while planning manufacturing, suppliers, and the broader ecosystem; experienced hardware operators remain especially valuable, although younger founders can pair with experienced staff. Unlike typical software, the managers say these companies often need hundreds of millions in first-round capital and substantial spending before a product exists; AI labs are reportedly signing with startups before hardware is available, and follow-on capital is available. Execution is constrained by permits, grid access, construction capacity, transformer and turbine shortages, and political headwinds, pushing some GPU-seeking companies toward Mexico, Australia, and other countries because of U.S. bottlenecks.
- Generalist, founded two years ago by former Google DeepMind researchers Pete Florence and Andy Zang, raised $200 million at a $3 billion valuation; it is pursuing a foundational “robot brain” designed to generalize across new physical tasks instead of requiring task-by-task training.
- Robot-foundation-model approaches are diverging: General Intuition is valued at $6 billion and converts video-game behavior into physical-world training data, with action labels capturing what players click and when as a claimed differentiator. Skilled AI’s S1 model claims that a robot can learn an untrained task from a single demonstration video, while the company emphasizes extensive pretraining and video-language models to reduce or eliminate post-training.
- Capital is rushing into robotics despite no consensus on the right technical approach; unlike autonomous vehicles, robots must physically manipulate and interact with objects, making valuable physical-world data a central open problem. Deployment remains early: companies are not operating at scale, current humanoid deployments are pilots in warehouses or factories, and safe interaction with humans—including robot alignment and potential workplace injuries—remains unresolved.
- Gatik raised $200 million while operating driverlessly on commercial fixed and dynamic routes; it has approximately $600 million in contracted revenue with customers including PepsiCo and Walmart and plans to use the capital to hire engineers and expand into additional cities.
- Open-source AI infrastructure is becoming an M&A theme: Hugging Face was reportedly nearing a $13 billion acquisition by Nvidia, although Nvidia had not confirmed the deal at recording time; Hugging Face had earlier rejected Nvidia’s reported $500 million investment amid concerns about single-investor control. The discussion argues that platforms enabling enterprise and developer adoption of open-source AI are attractive acquisition targets, suggesting further consolidation in this category.
- Persistent, collaborative agent infrastructure: The speakers described Grockbot as a cloud-hosted agent product that continues running while the user’s computer is off, unlike desktop-based OpenClaw and Hermes; they called it easier to use, though still requiring setup. They proposed a multiplayer, human-in-the-loop mode and argued that specialized agent swarms can build more context and expertise than a single agent.
- Enterprise software is becoming an agent interface layer: The Salesforce–Anthropic arrangement was described as placing Anthropic models inside Salesforce while allowing Claude to act as a front end over Salesforce data, workflows, and actions. The panel’s thesis is that systems of record remain the canonical read/write layer, while SaaS companies must optimize APIs, CLIs, and agent interfaces and accept externally created agents rather than trying to fully control the customer relationship. This is not a uniform SaaS outlook: mature horizontal systems of record were viewed as resilient because of enterprise compliance and accumulated reliability, while the durability of vertical SaaS workflows was questioned on a company-by-company basis.
- AI infrastructure demand and vertical integration: The podcast cited Nvidia revenue of $96.2 billion, 106% year-over-year growth, and 70% next-year growth guidance versus roughly 45% expected by Wall Street; it treated the figures as evidence that AI capex has longer-term momentum, while noting supply constraints. Speakers also discussed reported transactions involving Hugging Face at $12 billion and Poolside at roughly $6 billion as a push into open-source distribution, coding models, and agent tooling, but acknowledged that the Poolside announcement did not clearly specify whether Nvidia acquired its model, agent harness, or 100 engineers.
- Humanoid robotics’ key technical risk is physical-world generalization: The speakers said fixed competition behaviors such as running or jumping are comparatively tractable, whereas handling unprogrammed conditions and physical tasks—such as folding laundry or picking up objects without dropping or damaging them—remains difficult; they identified an Optimus-like robot’s ability to perform novel tasks as the critical test.
- Skilled AI’s S1 suggests a step change in robotics task acquisition. The model is described as a general-purpose robotic “brain” that can run on different robot bodies and learn tasks from a single human demonstration video, without task-specific programming, training, or teleoperation. It reportedly reduces a process that previously required 50–100 hours of puppeteering data and weeks of engineering to roughly 11 minutes of video ingestion; the system can adapt to substitute tools and retry after failures.
- Humanoid robotics is scaling rapidly, especially in China, but capability remains uneven. The Beijing humanoid games grew from about 280 to more than 600 teams, 500 to 2,000 robots, and 26 to 51 events year over year; the discussion cites first-half global humanoid shipments of roughly 19,000 units, up more than 270%, with China representing most shipments and buyers, alongside a reported RMB1 trillion state-backed humanoid fund. Robots reportedly surpassed the 9.58-second 100-meter record, but still struggled with mundane manipulation; force modulation, bimanual coordination, peg-in-hole insertion, and long-horizon recovery are presented as more meaningful tests of generalizability.
- Proprietary enterprise data is becoming a monetizable AI asset, with significant governance risk. Google reportedly bid about $10 million for Spirit Airlines’ digital estate in bankruptcy, above a $7.5 million bid from Meror, while judicial approval remained pending and employees questioned whether workplace communications could be sold as training data. The estate reportedly includes 30 million recorded customer-service calls, 15 million support-chat logs, and 7.5 billion passenger transaction records; the episode highlights the value of workflow data that captures how decisions and customer problems are resolved, while privacy and consent remain material risks.
- Physical-AI infrastructure: a16z’s American Dynamism team backs Hadrian and AMCA, which use advanced software and robotic systems to automate manufacturing and pursue price competitiveness with China. Training general-purpose robot brains, running simulated practice worlds, and incorporating field experience into fleet-wide model updates make data centers a foundational infrastructure layer for industrial AI.
- Company traction: Hadrian is reported to employ hundreds across three factories; its third facility in Mesa, Arizona, spans 290,000 square feet, represents a $200 million investment, and is expected to add more than 350 jobs, with production directed by AI.
- Market signal and constraints: Data-center construction is described as having surpassed a $50 billion annual rate, with a single large facility requiring up to 1,500 peak workers; the industry says it needs more than 300,000 electricians, while Microsoft’s president reportedly called the electrician shortage the biggest expansion obstacle. Projects perceived as intrusive, windowless boxes that reduce property values and create nighttime noise may face community protests, making design and local acceptance potential deployment constraints.
- “Quen 3.8 Flash Next” is presented as a next-generation open-weight mixture-of-experts model. Unlike the dense 27B version, it activates only part of the model per token, which suits systems with substantial memory and slower memory bandwidth; the presenter reports about 38 tokens per second on two DJX Sparks and says it can run on more modest hardware.
- Its cited architectural innovations are block-based QSA sparse attention for cheaper long-context processing, four-branch “gated residual” pathways that preserve information while other branches change, and “engram embedding,” which stores short token combinations in a lookup layer near the model’s start.
- The presenter claims the model already outperforms some leading open-weight systems and may rival the much larger “DeepSeek 4 Pro,” signaling continued rapid improvement in locally runnable, inference-efficient open models; the comparison is explicitly qualified as “maybe even.”
- AI infrastructure and industrial automation thesis: Garry Tan endorsed an essay arguing that data centers are foundational to AI-led manufacturing: robots require large-scale video and practice-run training, simulated environments, and fleet feedback loops, all of which depend on data-center compute.
- Company and market signal: The essay identifies a16z’s American Dynamism portfolio as backing Hadrian and AMCA, and describes Hadrian’s software- and AI-directed automated production across three factories; its Mesa facility is described as a 290,000-square-foot, $200 million investment expected to add more than 350 jobs.
- Buildout bottleneck: The essay reports data-center construction spending above a $50 billion annual rate, facilities requiring up to 1,500 peak workers, and industry demand for more than 300,000 electricians, framing skilled-trade shortages as a constraint on AI infrastructure expansion.
- Airbound’s founder Naman Pushp taught himself physics and engineering on YouTube during COVID and used that background to build drones; the company raised $37M from Greenoaks, Lachy Groom, and others while publishing a manifesto centered on a future where movement is in the air.
- Radiant Nuclear and Antares Nuclear were two of five companies selected for the U.S. Army’s $2.2B Janus program. Radiant is receiving $750M to build 15 Kaleidos microreactors, while Antares says its contract value is on the order of $1B and growing rapidly. Standard Nuclear also signed a binding agreement to supply Antares with TRISO fuel through 2035 for defense and space microreactors.
- Actinide became the first startup reported to enrich uranium and produce HALEU, an advanced-reactor fuel, using its Texas-built Endurance particle accelerator; the system had previously produced enriched Yb-176 for cancer-therapy applications.
- Anthropic previewed its Model Hardware Standard, designed to let AI agents operate multiple laboratory and manufacturing instruments—including microscopes, liquid handlers, and robotic arms—in parallel, for applications from drug-discovery experiments to quantum-computer laser calibration.
- Hugging Face’s Pollen Robotics, acquired by Hugging Face in 2025, unveiled Microduck, a $399 robot that users can train with reinforcement learning and that performs several locomotion and object-manipulation tasks out of the box.
- a16z’s American Dynamism team is described as backing Hadrian and AMCA to bring manufacturing back through advanced software and robotic automation; Hadrian’s thesis is that automation is necessary for U.S. production to compete on price with China.
- The essay highlights an embodied-AI infrastructure pattern: robots require substantial compute for video/practice-run training, general-purpose robot “brains,” simulated environments, and fleet feedback loops in which field data improves later versions.
- Hadrian is presented as deployment evidence for AI-native manufacturing: its machines run on software and production is directed by AI; the essay reports hundreds of employees across three factories and a 290,000-square-foot Mesa, Arizona plant representing a $200 million investment and more than 350 new jobs.
- Anthropic’s Fable 5 adoption appears constrained by retention requirements. Launched in June with a 30-day retention policy across its API, Bedrock, Foundry, Copilot, and coding tools, Fable 5 had no ZDR exceptions, and existing ZDR agreements did not apply to its traffic. Microsoft reportedly restricted employee use, GitHub disabled Fable 5 by default in Copilot, and Ramp usage data suggested traffic was routed away from it. Price was also a factor: Fable 5 cost $10 per million input tokens—roughly twice GPT-5.6—while Opus 5 was half its price; the author argues retention was the larger adoption drag, alongside price.
- Zero Data Retention is becoming an architectural differentiator and enterprise buying requirement. Standard APIs typically retain prompts and responses for about 30 days for abuse monitoring. ZDR removes logs, traces, and review queues after generation, except that harmful flagged data may be retained longer. Unlike a deletion policy that can be amended or overridden by legal process, architectural deletion leaves no customer data available to preserve. OpenAI’s Private Safety Processing preview keeps customer data on customer-controlled infrastructure or under customer-held keys while sending OpenAI only an automated misuse-alert category and severity.
- Privacy infrastructure and purpose-built data generation are emerging AI investment themes. The post expects ZDR passthrough, model portability, confidential inference, customer-held keys, and AI-traffic audit tooling to become more important as retention posture appears in security reviews. CISOs are also becoming more important participants in the buying process. Some high-value customers may pay to have their data forgotten, shifting model-improvement moats from customer traces toward RL environments, evaluations, and opt-in partnerships—the “RL factory.”
- Archil is expanding its infrastructure roadmap from a purpose-optimized file system toward a compute product: the company says its workloads now include post-training models running on bash, which it describes as creating major new demand for file systems.
- Archil is introducing persistent sandboxes on network-addressable compute for agent hypervisors with secure bash access, PostgreSQL databases, and Git remotes; it views these as a substrate for agents to build and manipulate stateful infrastructure.
David Ulevitch identifies AI and datacenters as prerequisites for U.S. reindustrialization and the jobs associated with it, signaling an investment theme around AI-enabled industrial capacity. A related post reports that Claude and Codex were important to setting up U.S. manufacturing in 2026, while Westmag’s MES, ERP, inline and end-of-line testing, and engineering stations used microcontrollers and tablets “vibe coded” by hardware engineers—evidence of coding agents entering factory engineering workflows.
- a16z launched the $1.1 billion Machine Age Fund for founders building the AI infrastructure and physical stack, including chips, memory, networking, systems software, power, and machines that bring AI into the physical world. Its thesis is that model improvements are outpacing memory, interconnect, power, and cooling, turning infrastructure bottlenecks into company-building opportunities and pushing strong teams from pure software into complex hardware.
- Capital intensity is becoming a competitive lever in AI: Ben Horowitz argues that a multibillion-dollar compute cluster can translate capital directly into capability and potentially erase software leads that would be difficult to overcome by simply hiring more engineers; he uses a $3 billion cluster and Grok as the example.
- a16z announced the new $1.1 billion Machine Age Fund for founders building across the AI infrastructure and physical stack, including chips, memory, networking, systems software, power, and machines that bring AI into the physical world.
- The fund’s thesis is that model progress is outpacing memory, interconnect, power, and cooling, turning infrastructure constraints into company-building opportunities; a16z says strong teams are moving from pure software into complex hardware.
- a16z is deliberately framing the opportunity as “machine intelligence,” arguing that current systems learn from accumulated human knowledge and that capital deployed into models is increasingly constrained by the underlying machines.
- A wearable-input product concept is emerging: use inexpensive Bluetooth rings as physical triggers for voice transcription tools such as WisprFlow.
-
The thread links to the
remotehumans/riffGitHub repository as a related project.
- a16z announced the $1.1 billion Machine Age Fund to back founders rebuilding the AI stack across chips, memory, networking, systems software, power, and machines that bring AI into the physical world.
- The fund’s thesis is that AI bottlenecks have shifted below the model layer: model progress is outpacing memory, interconnect, power, and cooling capacity, creating company-building opportunities in complex hardware and the broader physical stack.
- a16z partners cited strong infrastructure demand and scarcity: hyperscaler capital expenditure is described as roughly $700 billion this year and potentially $1 trillion next year, while supply is reportedly booked through 2028 and some GPU lots have drawn multi-day auctions.
- Martin Casado identifies chips, memory, interconnects, storage, and robotics as parts of a compute-hardware infrastructure sector undergoing what he calls its largest transformation in 30 years; he says computer science will be central to addressing the challenges and that “we’ve raised $1.1B to help that along.” The post does not specify the company, financing stage, or lead investors.
- AI infrastructure investment thesis: a16z announced the $1.1 billion Machine Age Fund to back founders building across chips, memory, networking, systems software, power, and machines that bring AI into the physical world. The thesis is that model progress is outpacing the memory, interconnect, power, and cooling infrastructure beneath it, making complex hardware constraints—and teams moving from pure software into hardware—major company-building opportunities.
- Shift toward autonomous AI employees: Martin Casado describes AI products evolving from search and chat interfaces toward agents that have their own computer and browser and can execute high-level tasks like email triage, while checking with the user before taking consequential action.
- a16z announced the $1.1 billion Machine Age Fund to back companies rebuilding the AI physical stack, including chips, memory, networking, storage, systems, data centers, robotics, and home AI appliances; its thesis is that supply-chain, physics, and computer-science constraints require infrastructure innovation down to electricity.
- The fund highlights a severe infrastructure supply-demand gap: compute density per rack has increased 28× from H100 to Rubin; rack power has risen from roughly 5–10 kW to 100–250 kW and is expected to reach 1 MW within three years; data centers are expanding from tens to hundreds of megawatts and, in some cases, gigawatt-scale campuses, while hardware supply has historically grown only 20–30% annually versus the triple-digit growth a16z says demand requires.
- a16z says hardware startups now represent more than 20% of its deal flow and is making hardware an official investment motion. Its relevant bench includes former Intel Data Center Group CTO Guido Appenzeller, longtime data-center executives Raghu Raghuram and Martin Casado, AI-infrastructure investors Shangda Xu and David George, and American Dynamism investors David Ulevitch and Erin Price-Wright; recent investments include Unconventional AI, Nexthop, Volta, Atoms, and Mind Robotics.