Winning the AI Race Part 1: Michael Kratsios, Kelly Loeffler, Shyam Sankar, Chris Power
Summary
Washington’s 90-action AI plan treats innovation, physical infrastructure and global ecosystem reach as one national-security strategy. Jacob Helberg laid out the plan’s pillars: America must out-innovate competitors, accelerate data centers, energy and domestic manufacturing, then create the AI stack for the world. The plan targets actions achievable within six to nine months because “you can’t regulate your way to winning the AI race.”
The investable bottleneck is shifting from model intelligence to power, permitting, machine tools and skilled labor. Michael Kratsios wants federal scientific data made usable, not merely open, while warning that AI will enter regulated products from drones to medical diagnostics. Jacob Helberg separately flagged more than 1,000 proposed or enacted state measures as a potential path toward “a patchwork of 50 different state regulatory regimes.”
Hadrian’s factory results support the episode’s central labor thesis: AI can create industrial capacity where qualified workers no longer exist. Chris Power reported 4x manufacturing productivity, 10x workforce productivity and 30-day training for recruits entirely from non-factory backgrounds; the company’s Arizona expansion is planned at four times the Los Angeles facility’s size with 350-plus jobs. “Hadrian’s advanced factories look and operate more like a data center.”
Gecko Robotics reframed energy as both AI’s constraint and one of its highest-return applications. Jake Loosararian’s example began with a 620-megawatt plant producing only 580; robotic inspection and AI reportedly unlocked 1% efficiency, which he extrapolated to 11.9 gigawatts across the US thermal fleet “without putting a shovel in the ground.” His call: “AI shouldn’t just consume, it should create energy.”
Palantir’s Shyam Sankar sees the strongest adoption where frontline workers author the applications and institutional leadership releases their agency. He wants workers made 50 times, not merely 50%, more productive; cited factory training falling from three years to three months; and described a four-week fellowship for mechanically intuitive, often self-taught workers. His categorical conclusion was that “the traditional college degree is dead.”
Paul Buchheit’s abundance case is that natural language expands the pool of builders, while capital intensity preserves scarcity at the model layer. With only 2%-3% of Americans able to code—and perhaps half that number capable of building a startup—“English is the new programming language” could produce 10x or 100x more startups, robotics companies and local applications. Buchheit expects the number of foundation-model providers to remain relatively stable, with open source constraining censorship and lock-in among closed vendors.
Small businesses are positioned as AI’s distribution channel, but the resulting market may be a barbell rather than a universal uplift. Kelly Loeffler said 60% of the SBA’s $21 billion lent this year went to companies with one to five employees; Keith Rabois expects those firms to gain incumbent-grade information, products and administration, then take share from the mid-market while compute leaders such as NVIDIA also benefit. The counterweights are energy and materials costs, industrial supply-chain exposure, rigorous underwriting and Power’s warning that offshore competition remains “companies versus the CCP.”
Deep dive
1. America’s AI plan couples software leadership to physical capacity
Helberg reduced the administration’s plan to three linked pillars: innovation, infrastructure and ecosystem. The private sector must out-innovate global competitors; America must build the data centers, power generation and manufacturing supporting AI; and the United States must create the AI stack for the entire world. “There’s just no substitute for innovation.”
The document contains 90 executive-branch actions, deliberately framed as an action plan rather than another long-range national strategy. Kratsios said its authors concentrated on measures achievable within six to nine months, reflecting the view that AI’s economic and national-security consequences make speed itself a competitive asset.
The process began with a public request for information that drew more than 10,000 responses, including submissions from technology companies and Hollywood actors. Kratsios took the breadth as evidence that AI already touches nearly every American industry; Helberg emphasized the same economic arc, from newly possible AI businesses to enabling sectors such as mining, energy, chips, fabs and data centers.
2. Regulation must follow AI into industries without fragmenting the market
Kratsios said there will not be one statute comprehensively “regulating AI.” Models will instead become components of drones, autonomous vehicles and FDA-approved medical diagnostics, putting responsibility inside existing sectoral agencies. His standard was straightforward: those agencies should protect their domains while creating conditions in which AI-powered technologies “can thrive and not be hindered by the government.”
Government data is another innovation lever, but Kratsios distinguished nominal openness from usefulness. Department of Energy laboratories hold datasets potentially valuable to materials science and medicine, yet “dirty, nasty data that isn’t homogenized” will not train much. He cited $150 million allocated to DOE for an AI-for-science program focused partly on making such assets usable.
On talent, the discussion moved beyond elite researchers. Kratsios endorsed attracting leading scientists and engineers, then stressed shortages among electricians, HVAC technicians and the other trades needed to construct AI infrastructure. The action plan therefore directs labor and reskilling programs toward the workforce behind the data centers, not only the people designing their models.
Federal preemption produced a meaningful split. Kratsios declined to make it central because the plan covers actions the executive branch can take without Congress; Helberg nevertheless called more than 1,000 proposed or passed state AI measures a brewing national-security threat. Fifty separate regimes, he argued, could hobble a national network while China treats AI as a unified strategic priority.
3. Industrial decline is presented as the binding national-security risk
Power’s historical chain was blunt: industrial leadership produces military capacity, military capacity helps sustain reserve-currency power, and successful countries eventually become complacent enough to offshore their industrial base. America won World War II because commercial producers could pivot—Ford from cars to bombers, watchmakers to naval instruments—not because every weapon was individually superior. “Our tanks weren’t so great. We just had tons of them.”
His comparison with China was designed to make the capacity gap visceral. Power said Chinese munitions factories can produce roughly 1,000 units annually, while US war games exhaust missiles in seven days and replacement can take three years; Chinese shipbuilding capacity is 200 times America’s, while the US produced five ships in the cited year.
The vulnerability extends through pharmaceuticals, drones, iPhones and production equipment, but Power’s most urgent concern was human capital. America retained an aging cohort of patriotic, highly skilled tradespeople supporting defense suppliers, while manufacturing depth moved offshore. “You could give me a billion dollars and we can’t hire” the millions of welders and machinists required, he argued.
4. Hadrian uses automation to multiply scarce workers, not eliminate them
Hadrian’s development path began by operating a factory while simultaneously building its software, followed by roughly 18 months of beta work with aerospace customers. Power said Factory 2 launched in 2024, scaled the company 10x in one year and now produces micron-tolerance components and larger assemblies for rockets, satellites, aircraft and drones through its Opus autonomy platform.
The operating benchmarks carry the economic argument. Power put typical US factory uptime at only 20%; Hadrian claims a 4x jump in manufacturing productivity and 10x improvement in workforce productivity. That leverage is essential, in his telling, because the labor pool is too small to rebuild ships, submarines, munitions and drones through conventional staffing.
Training time fell to 30 days, and Power said 100% of Hadrian’s workforce came from outside factories: high-school graduates, veterans, nurses, bus drivers and former desk workers. One employee hired from stocking shelves at Home Depot progressed to running 10 machines simultaneously; exposure to software and AI has also moved recruits into management and software-engineering roles.
The next Arizona factory is planned to open within six months, around Christmas, at roughly four times the size of the Los Angeles operation and with more than 350 new jobs. Hadrian then intends to launch specialized facilities for ships, submarines and munitions in 2026, ultimately pursuing factories in every state.
5. Reshoring economics turn on energy, trade policy and machine tools
Power separated products America must reshore for security—submarines, ships, munitions, rare-earth magnets and drones—from commercial volume that can still be sourced offshore. Hadrian can compete in the onshore-required defense and aerospace market today, he said, but winning the offshore market, which he described as 10 times larger, requires tariffs because Chinese competitors benefit from nationally subsidized energy and materials.
When pressed on whether Hadrian needed direct government support, Power’s answer narrowed the ask: defense economics already work, while commercial reshoring needs a level playing field. Energy is the “silver bullet” because it drives both factory operations and raw-material inputs; he claimed energy represents roughly 90% of the cost embodied in aluminum and steel.
There is also a supply risk in “the machines that build the machines.” Power said America invented many advanced machines through the Air Force but forgot how to make them; Hadrian avoids Chinese equipment over cybersecurity concerns and relies chiefly on suppliers in Germany, South Korea and Japan. His technical bet is that these machines are “really dumb computers” whose performance American software and AI can amplify.
Manufacturing’s missing digital history forced Hadrian to build scheduling, control and modeling systems itself. Customers still transmit requirements through 20-page PDFs “full of hieroglyphics” that can consume 50 expert hours; vision models parse them, while production stays about 80% automated with humans labeling outcomes. Reinforcement learning then links process settings to whether the finished part actually passed quality checks.
6. Physical data can make AI a producer of power
Loosararian said Gecko Robotics now manages data concerning more than 500,000 critical infrastructure assets across energy, mining, metals, manufacturing and defense. The obstacle is not executive interest in AI but missing inputs: crucial information is still gathered manually by “Joe,” using century-old tools in dangerous environments, leaving software engineers little structured physical data to model.
Gecko fills that gap with flying, swimming, crawling, walking and wall-climbing robots. At an operating gas plant, robot dogs collect operational readings while other devices map asset health; the data feeds Cantilever, Gecko’s AI-powered operations platform. The organizing principle is to build the software and its ontology upward from first-principles measurements of the actual asset.
His concrete example was a plant rated for 620 megawatts but running at 580. Combining robotic health data with the customer’s operating records identified a steam problem entering the turbine; Loosararian said fixes like this have yielded approximately 1% efficiency improvement. Applied across the US thermal fleet, he estimated 11.9 gigawatts of additional power without new construction.
Longevity may matter as much as efficiency. Citing a DOE study, Loosararian warned that grid assets face roughly four years of remaining useful life and that uncorrected trends could produce 100 times as many blackouts by 2030. Gecko’s data has reportedly supported extensions of 10, 20 or 30 years; he said the average across the assets discussed has been about 35 years. “The physical world has been forgotten about,” but AI can expose capacity hidden inside it.
7. Frontline deployments remove paperwork while preserving judgment
At Tampa General, nurse Laura Deirdinas described assembling reports across 32 neurointensive-care patients from charts, conversations and stapled paper that could already be outdated. AI pre-assembled the information for multidisciplinary rounds, enabling a charge nurse to report confidently and returning time to the bedside: “We are the heart of healthcare.”
At submarine supplier PRL Industries, engineers had spent three days assembling quotes from paper archives, tables, email and side communications. Its AI tool now gets halfway through that workflow in minutes and tracks each part’s exact status, allowing management to prioritize components holding up ship construction. The claimed payoff is matching quality-management speed to the workforce and machines.
Julie Nordberg’s rural Michigan hospital is four hours from the next comparable facility and lacks the manpower of an academic center. A shared AI-supported facility snapshot reduced chart review and meetings, while earlier detection could trigger earlier treatment. Her boundary was explicit: the system “could never replace our frontline staff”; its value lies in background legwork.
Panasonic Energy’s example paired historical maintenance records with live machine data to detect failure warnings and dispatch technicians before breakdowns. The site has produced more than 11 billion batteries in eight years and recruits from tourism, hospitality, automotive work and even slot-machine repair; an interactive learning tool reduced lost confidence and turned frontline supervisors into product designers.
8. Worker agency determines whether enterprise AI compounds
Sankar proposed children’s enthusiasm as the ultimate adoption test: one featured worker brought her daughter, and another described how AI had changed her view of her children’s future. His ambition is not a 50% efficiency gain but making the American worker “50 times more productive,” combining US model leadership with frontline ingenuity.
The adoption divide, in his experience, is institutional rather than sectoral. Organizations win when they liberate workers to design applications; top-down force-fitting suppresses the mechanical intuition of people closest to the process. AI also increases the value of exceptional human judgment: it makes “the person with the greatest taste more valuable” and distributes that taste across the organization.
Sankar expects a high “cardinality of agents and models.” General systems provide the starting point, but companies capture additional alpha by specializing models around what makes their operations different. Healthcare illustrates the opportunity: forced electronic-health-record adoption roughly halved patients seen per hour, so AI should be designed backward from care delivery and reduce time spent facing computers.
Reindustrialization offers the largest prize because coordination creates so much “dwell time.” A submarine supplier waiting for data or approval leaves expensive equipment idle, and every supply chain is constrained by its weakest participant. Sankar cited Panasonic training falling from three years to three months, then described Palantir’s four-week American Tech Fellows boot camp for overlooked, mechanically gifted autodidacts. “The traditional college degree is dead.”
9. Natural language expands the startup funnel and the problems founders attempt
Buchheit’s premise was that “English is the new programming language.” Only 2%-3% of Americans know how to code, he estimated, and perhaps half of those can do it well enough to found a software company. AI therefore extends Y Combinator’s original 20-year thesis—that two people living cheaply can start a company—toward a plausible 10x or 100x expansion in founders.
He expects creation to become an iterative conversation: describe the desired product, inspect the result, then say, “not quite like that, more like this.” The social objective is dispersing “tools of wealth creation in as many hands as possible,” including people building useful applications for one town or community rather than chasing the next Google.
Physical AI is already widening the ambition set: Buchheit found the number of robot arms at the latest YC Demo Day striking. As software tasks become easier, founders tackle harder physical problems. The earlier no-code wave was a false start but revealed the latent supply of builders whose strengths are design, language or emotional intelligence rather than formal computer science.
10. Intelligence abundance opens science and media while model supply stays scarce
In a separate host-led extrapolation, the host reduced wealth creation to two inputs, energy and intelligence, then predicted that a 10x increase in global intelligence could enable a corresponding 10x increase in wealth. The host also predicted that AI science labs could generate their own experimental data and, within 20 years, models could predict how drugs affect the body without testing and become more predictive than today’s clinical trials.
The hosts argued that generative media could democratize production from another direction. A child in middle America with a vision for a Disney-style movie may soon produce it without a $100 million budget, replacing a small number of elite production jobs with many more creators whose communities, countries and sensibilities are currently absent from mainstream media.
At the foundation layer, however, Buchheit expects a relatively stable provider count because training costs are astronomical. He wants multiple closed and open choices: open models keep proprietary vendors “honest” by giving users somewhere to go if capabilities are disabled or speech is excessively constrained. On OpenAI, he said open source was never specifically promised; a host pushed back that its origins implied more openness.
Buchheit recalled OpenAI beginning at YC in 2015 partly because advanced AI appeared locked inside Google. The same platform risk now confronts Facebook: his children spend time talking to AI rather than using social media, making ChatGPT a direct competitor for attention previously captured by Instagram. Character-driven systems suggest personality may become as strategically important as raw model capability.
11. Small businesses are positioned to transmit AI into the real economy
Loeffler’s showcase was a 60-person bicycle factory in Seymour, Indiana, reviving an industry she said had become 98% import-dependent over 30 years. AI and advanced manufacturing make such facilities viable in towns needing jobs; repeat that pattern across strategically important goods, she argued, and AI becomes “a job creation machine for reshoring.”
She placed the labor shortage at 7-12 million open US jobs, concentrated heavily among small businesses, while pointing to an asserted $15 trillion investment pipeline. Her historical comparison was 56 million workers in 1940 versus 170 million today: only 40% of 1940’s jobs still exist, while she attributed 85% of today’s jobs to technological advances.
Scale makes the channel consequential: Loeffler cited 34 million small businesses against only 20,000 large companies. That includes manufacturers with hundreds or even 1,500 workers, but also prospective “solopreneurs” building software or operating advanced equipment with tiny teams. Her phrase for the cultural reset was “Main Street is going mainstream.”
Rabois supplied the competitive mechanism: AI gives small firms incumbent-level research, marketing, legal and accounting knowledge; lets an HVAC contractor offer an application comparable to a large commerce platform; and reduces administrative expense, including 5%-15% waste identifiable through auditing. He expects a barbell—compute leaders such as NVIDIA benefit, while newly capable small firms pressure the mid-market.
12. SBA expansion is constrained by underwriting, not ambition
Loeffler said she reversed a prior SBA restriction preventing government-backed loan proceeds from purchasing AI technology, opening financing for implementation, CNC equipment, advanced manufacturing and training. Of $21 billion lent so far that year, 60% went to businesses with one to five employees; the agency was processing about 2,000 Main Street loans each week.
The SBA does not lend directly: thousands of banks originate loans backed by a government guarantee, while the Small Business Investment Company program supplies an equity-oriented channel and previously backed Tesla. Loeffler said the agency returned staffing and spending to pre-pandemic levels, brought employees back to offices or field locations, and was nevertheless on pace for a record lending year.
Risk discipline was non-negotiable. Loeffler said looser underwriting had helped drive portfolio losses up by $400 million; her target loss ratio is 3% or less, which allows the core program to operate without taxpayer subsidy. She also rejected relaxing standards indiscriminately for favored sectors: “We can’t put some small businesses on the hook for other small businesses.”
Defense, medical devices, minerals and other critical industries may still justify different loan or equity structures, and Loeffler said discussions with the Defense Department and experiments inside SBIC were under way. A host argued that 10% equity in a success such as Tesla could have covered “a hundred Solyndras”; another host warned of adverse selection. The unresolved design problem is how to give taxpayers potential upside without abandoning prudent underwriting or having government pick winners and losers.