Erin Price-Wright

America’s AI constraint is increasingly “atoms and not algorithms”: minerals, power conversion, factories, and transmission must scale alongside models and chips. Turner Caldwell says the US is “50 years behind” China on critical-mineral supply, while Drew Baglino sees a century-old grid that remains largely mechanical, overbuilt, fragile, and dependent on overseas suppliers.

Permitting alone cannot close the minerals gap because execution after approval is still painfully slow. A project can take five years to build and another three to five years to reach operating rate; Mariana Minerals therefore targets design, construction, procurement, and ramp-up, aiming to build 10 projects in 10 years.

Both founders are embedding software in physical-asset operations rather than merely selling standalone software. Mariana uses agentic workflows and reinforcement learning across mines and refineries, including an effort to “remove humans from the loop,” while Heron Power uses silicon and software to replace “steel, oil, and copper” in grid-scale power conversion.

US manufacturing’s central disadvantage is supply-chain geography, not factory wages. Baglino estimates labor represents less than 10%—and potentially less than 5%—of cost of goods sold in a modern automated factory; China’s advantage is that the roughly 7,000 parts needed for a car can sit within a three-hour drive.

The Tesla operating model offers industrial startups a combination incumbents struggle to reproduce: techno-optimism, risk tolerance, mission, and persistence. Caldwell’s sharpest formulation was that Tesla keeps “barreling through the challenges as long as the outcome is worth it”; Baglino added the concentrating effect of knowing “whether or not the paycheck will clear” depends on execution.

The requested policy is durability and planning certainty. Caldwell wants the minerals equivalent of the tools built for oil and gas over the past 50 years; Baglino wants manufacturing zones, aligned jurisdictions, and a “federal highway trust fund for the grid” so private capital and suppliers can plan confidently.

🔗 Original source & video: The Founders Who Left Tesla to Rebuild America | a16z

Tesla and SpaceX compress hardware timelines by pairing flat information flow with decisive technical leadership. Junior engineers can reach decision-makers directly, while leaders absorb the risk of imperfect information and say, “Go.” The operating loop is to gather as much evidence as the deadline permits, place the bet, test it, and iterate.

Aggressive milestones are diagnostic tools for exposing the few constraints that actually govern schedule. Price-Wright’s framing, endorsed by Caldwell: if 1,000 things must happen and 900 fit inside six months, the team should attack—or delete—the 100 that do not. Chandler Luzsicza applies that logic to Galvadyne’s goal of getting a rocket airborne by June.

Long hours are not the primary cause of burnout; organizational churn is. Luzsicza argues that politics, erratic priorities, data silos, and teams “hoarding your Legos” destroy motivation by separating effort from visible progress. Impossible goals can energize a mission-aligned team, but only when a credible technical path makes them aggressive rather than imaginary.

The factory mindset extends beyond production lines into design, laboratories, construction, refineries, and mines. Luzsicza’s Starship lesson is to question requirements until “simple is fast, simple is cheap”; Caldwell’s is to give every activity a takt-time analysis and measurable daily or hourly output. Mariana Minerals sees the missing software backbone for that control layer as the opportunity in major players that are 50 to 100 years old.

Vertical integration should clear an existential test, not merely promise lower component costs. Caldwell would not integrate early to save 5%, 10%, 20%, or even 50% unless the company otherwise cannot exist; Luzsicza prioritizes assemblies that would bottleneck Galvadyne’s path toward 10,000 missiles per year. Integration also internalizes the supplier’s operational risk and upstream supply chain.

Talent density is manufactured through unusually deep technical screening and extended trial periods. Tesla candidates may face six engineers, a technical test, and eight to 10 conversations; SpaceX’s internships provide a three-month proof period that repeatedly converts into critical full-time talent. Caldwell, who interned there four times, calls that funnel “so freaking crucial.”

Future founders should accumulate complete execution cycles before trading technical learning for company-building risk. Caldwell recommends seeing projects through the early, middle, and deployment “messy phases” multiple times, learning “what good looks like,” and building credibility to recruit exceptional people. His closing hierarchy is clear: over-index on technical depth before learning fundraising and hiring; Luzsicza agrees that founders should not try to learn how to build rockets on the job.

🔗 Original source & video: What Tesla and SpaceX Teach Founders About Building Hardware | a16z

GPT-5 is less a frontier-compute leap than an “economic release” built around routing. Dylan Patel argues that power users lost access to GPT-4.5 and o3—with o3 thinking roughly 30 seconds on average versus GPT-5’s 5-10 seconds—while free users sometimes receive reasoning they never had before. The router lets OpenAI choose regular, mini, or thinking models and “gracefully degrade” service, trading maximum capability for dramatically greater token capacity.

The router’s larger prize is matching inference spend to each query’s monetizable value. A “why is the sky blue?” request can go to mini, while a search for the best DUI lawyer, flight, or product can receive “ungodly amounts of compute” because OpenAI could complete the transaction and take a cut. With an estimated 10% of Etsy traffic already coming from ChatGPT, Dylan sees agentic commerce—not result-degrading ads—as the route to monetizing free users.

Flat-rate AI subscriptions are colliding with 20x differences in consumer usage. Anthropic and coding products have tightened rate limits after heavy users exploited “negative gross margin” plans; one developer reportedly rearranged sleep into sailors’ power naps, while a Reddit leaderboard included usage worth roughly $30,000 a month. Enterprises may support commitments or averaged flat fees, but consumers increasingly point toward usage pricing—even as products use subscriptions and superior review interfaces to create stickiness.

AI may already create more value than its infrastructure costs, but the labs capture only a fraction of it. Dylan Patel’s coding thought experiment—30 million developers, productivity doubled, $100,000 of value each—produces $3 trillion of potential GDP value from one use case, while Dylan believes OpenAI captures “not even 10%” of the value ChatGPT has created. That mismatch need not stop capex: hyperscalers could grow spending another 20-30%, with CoreWeave, Oracle, infrastructure funds, and sovereign capital adding less immediately economic capacity.

Custom silicon is Nvidia’s largest threat only if AI demand remains concentrated among a few giant buyers. Google is making millions of highly utilized TPUs, Amazon millions of Trainium chips, and Meta is sharply increasing internal-silicon orders; Dylan thinks Google should physically sell TPUs, not merely rent them. If open models and cheap deployment disperse demand, however, Nvidia’s universal ecosystem strengthens—and independent challengers must be “like 5x better” before supply-chain, software, and margin disadvantages erase the lead.

American AI deployment is constrained less by electricity’s cost than by powered sites, grid equipment, and construction speed. Dylan puts roughly 80% of a Blackwell data center’s cost in GPUs, networking, buildings, and power-conversion capital, leaving only 20% for land, electricity, cooling, backup power, and related items. That makes paying extra to launch three months earlier rational; meanwhile, China’s current constraint is capital and chip quality rather than power, despite its ability to scale generation faster.

Intel needs immediate operational surgery and capital, while several platform incumbents need product urgency. Dylan says Intel’s five-to-six-year design cycles and as many as 14 silicon revisions must fall toward two-to-three years and one-to-three revisions; without a major cash infusion or severe cost cuts, it could “literally” go bankrupt before a formal separation is completed. His broader calls: Nvidia should reinvest its projected $100 billion-plus cash pile into infrastructure, Google should open TPUs, Apple should spend perhaps $50 billion on AI infrastructure, and Erik says Microsoft must “shake the crap out of the company” despite its extraordinary starting position.

🔗 Original source & video: Dylan Patel on GPT-5’s Router Moment, GPUs vs TPUs, Monetization

Critical minerals are physical inputs to AI, grids, batteries, data-center infrastructure, vehicles, and defense systems. Turner Caldwell’s mass-flow call is emphatic: “We need a lot of aluminum. We need an insane amount of copper. We need more iron. We need more zinc,” while lithium production capacity must roughly 4× over the next 10 years if planned batteries are to be built.

A mineral project is a site-specific chain from sub-1% ore to high-purity metal, not a fungible factory template. Ore bodies change grade and impurity mix as mining proceeds—“the Earth is heterogeneous”—so every flowsheet is bespoke and needs flexibility. That makes recovery and adaptability economically decisive because every lost atom must be mined again.

The venture thesis is to own the mine-to-refinery operating system because selling point technology into incumbent miners has become a “death spiral.” Multi-billion-dollar plants resist changes that might create multimillion-dollar downtime, pilots can miss commercial builds that arrive perhaps once every five years, and operators distrust outsiders touching their “cash register.” Mariana, emerging with $85 million raised, is betting vertical integration can capture efficiencies that point-solution SaaS vendors cannot capture on their own—but it also imports the partner’s risk into its expanded risk profile.

Mariana’s software bet targets two large sources of wasted time: a roughly three-week construction-information lag and refinery-control problems with about 1,000 interacting variables. Capital Project OS would automate engineering and procurement workflows; Plant OS would use reinforcement learning to optimize refinery operations, including recovery, energy, and reagents, across 24–48-hour feedback loops. Caldwell wants to remove humans from many operating decisions.

China’s moat is skilled execution capacity as much as policy or capital. Caldwell saw 13,000 people mobilized at a Chinese-backed Indonesian nickel refinery during construction and commissioning; a U.S. project might struggle to field one-tenth as many. Indonesia now supplies “something like 70%” of global nickel, illustrating how labor depth and downstream buildout compound geopolitical leverage.

The portfolio call is countercyclical: diversify, navigate frothy markets, and build at commodity troughs. In the rare-earth discussion, Caldwell calls the market “a little bit of a frothy market,” while focusing on lithium and regarding copper demand and declining grades as “pretty hard to ignore.” The operating bet is that software-controlled circuits can process lower-grade copper without meaningful cost inflation.

U.S. supply security needs faster exploration, permitting, and demand support. Exploration over more than five acres on federal land can require BLM approval, while price floors or fixed-price offtakes—like the cited MP Materials arrangement—could unlock infrastructure capital that will not underwrite commodity volatility. Mariana’s mission is concrete but hedged: build 10 increasingly large projects in 10 years, expand overseas and perhaps underwater, and restore confidence that complex minerals infrastructure can be built “cost-effectively, time-effectively and responsibly.”

🔗 Original source & video: The U.S. Can’t Build AI Without These Materials

America’s grid problem is institutional atrophy as much as physical capacity. Ryan McEntush says the system “effectively froze” in the early 2000s as manufacturing shifted to Asia, leaving operators without the workforce or skill to plan and execute large projects quickly or cheaply. David Ulevitch adds the stark comparison: U.S. per-capita energy use peaked in 1973, while China’s energy use increased ninefold over the same period.

Distributed generation and storage could leapfrog a grid where interconnection can take a decade and transformer backlogs reportedly exceed 20 years. David describes the grid as roughly century-old technology operating near capacity; Erin Price-Wright argues for placing solar, batteries and load together, bypassing some interconnection and delivery infrastructure whose costs keep rising even as generation gets cheaper. Data centers already embody the urgency: “I can’t afford to wait 10 years… I need this power now, today.”

Texas is the panel’s proof point for deploying cheap solar and batteries at speed. Erin says the state roughly doubled solar capacity in three years and added thousands of batteries, improving its ability to absorb sharp demand changes; Ryan argues every state should study ERCOT’s distributed model, while acknowledging that regulated markets make replication harder. Erin describes the manufacturing dependence on China, while David warns that losing Chinese battery supply “could be catastrophic… in a very, very short period of time.”

The winning energy mix is “yes, and,” not a single-resource bet. Erin’s personal bet is that solar and batteries will remain cheap and fast, but she says gas, nuclear, geothermal and hydro will remain necessary because the long-tail cost of intermittency becomes severe once variable resources reach roughly 50%-75% of the grid. Her broader point is about load design: expanding data centers, EVs, heat pumps and autonomy will enlarge both baseload and daily peaks, so the system must not oversolve for either.

Flexible compute is a more plausible demand-response asset than centrally controlling Americans’ thermostats. David says consumers will “flatly reject” dictated indoor temperatures, while Erik points to crypto mining and David suggests shifting noncritical data-center jobs to cheaper hours. David also pushes back on Erin’s view that data-center demand may be overstated, arguing society may instead be underestimating AI electricity consumption across the next 10-50 years.

Grid software is the clearest venture-scale layer in an otherwise capital-heavy rebuild. Erin explains that the grid lacks the internet’s bidirectional communication, data layer and control plane; David argues that software should enter from the edges rather than through a slow top-down utility rollout. The grid lacks “no Splunk,” “no Palo Alto Networks” and “no Looker,” while bottom-up telemetry from batteries, chargers and distributed generation could improve forecasting that still relies heavily on weather. Erin says AI could also turn permitting work performed by armies of consultants over months or years into “minutes or hours.”

Nuclear combines baseload, resilience and defense value, but scaling it requires rebuilding America’s megaproject muscle. David highlights Radiant Nuclear’s proposed 1-megawatt, truck-transportable microreactor and separately imagines transportable reactors being flown on C-130s. He contrasts that flexibility with military fuel that can cost more than $200—and sometimes $400—per gallon to deliver. The closing priorities are categorical: Ryan says there is no safety, national defense or national security without a reliable electrical grid; Erik says energy policy should prioritize power that is “cheap, reliable and clean—in that order.”

🔗 Original source & video: America’s Energy Problem: We Need A New Grid