Marc Andreessen - The Battle For Tech Supremacy - [Invest Like the Best, EP.410]
Summary
DeepSeek R1 resets AI’s cost curve by making frontier reasoning open, self-hostable, and roughly 30X cheaper. Marc says the full model can run on about $6,000 of hardware, while distilled versions already run on MacBooks and iPhones; V3 and R1 also ship with code and technical papers that become a roadmap for competitors. His categorical forecast: within five years, everyone will have a “superhuman AI lawyer” and “AI doctor” on their phone.
The immediate winners are AI users, while proprietary model labs and NVIDIA face price pressure—but Jevons Paradox could ultimately lift the entire stack. OpenAI and Anthropic face a free, open alternative, while DeepSeek’s efficient use of cheaper chips challenges NVIDIA’s margins of as much as 90%. Yet a 30X price decline might drive 30X, 100X, or 1,000X more usage: reasoning is “fire being brought down from the mountain,” and demand could outrun falling unit costs.
DeepSeek turns U.S.–China competition into a contest over which country can offer the world the more open technology stack. China’s “judo move” is that an ostensibly closed system now offers the best open-source system while American companies called OpenAI and Anthropic remain closed. Because AI will enter healthcare, education, transportation, telecom, energy, and national infrastructure, Marc argues the West must answer by returning to “freedom and democracy and openness,” not by adopting Chinese-style control at home.
Attempts to contain China can force exactly the innovations and competitors U.S. policy seeks to prevent. Chip restrictions helped motivate DeepSeek to optimize cheaper hardware, encourage a parallel Chinese semiconductor industry, and may accelerate the timeline for a Chinese invasion of Taiwan because of its “magic fabs.” Copyright presents the same dilemma: if Chinese models may train on every book while U.S. models are prohibited, Marc warns that “they win and we lose,” though DeepSeek’s actual training corpus remains undisclosed.
AI should commoditize investment analysis, but Marc doubts it will soon replace early-stage venture’s interpersonal “alchemy.” Models such as o1, o3, o7, or R4 will probably read annual reports better than most humans, but that advantage becomes standard once every investor adopts it. Founder evaluation, coaching, and “project picking” remain face-to-face judgments developed over years; Marc’s hedge is refreshingly falsifiable: “If I meet the algorithm that can do that better than I can, I will instantly retire.”
Andreessen Horowitz is being built as an enduring operating company while keeping early-stage venture non-negotiably at its core. The firm pairs timeless principles—respect founders, know the details, avoid “drive-bys”—with rapid adaptation, including a major political operation that did not exist four years earlier. It rejects becoming an asset gatherer harvesting fees: a $5 million Series A can offer as much return opportunity and strategic learning as a $500 million later-stage investment.
Politics has become the fourth—and recently the dominant—constraint on technological returns. Marc’s original framework aligned technology, social readiness, and capital markets; government now overlays all three, and he describes the past four years’ treatment of crypto, fintech, social media, and AI as the largest bottleneck. His broader power thesis is that collapsing institutional trust produces counter-elites: “a small organized elite” always governs a large disorganized mass until a rival group persuades that mass it can represent them better.
China’s phones-to-drones-to-cars-to-robots supply-chain ladder makes physical technology the next geopolitical pressure point. China holds “ninety-nine point something percent” of drones, Unitree sells a $1,500 robot dog against Boston Dynamics’ $50,000–$100,000 equivalent, and humanoids might arrive at $10,000–$20,000—extraordinary abundance with embedded security risk. In defense, one planner told Marc that backpack suicide drones are the biggest innovation since the stirrup: 20 trained soldiers with drones might hold off thousands or tens of thousands of conventional troops.
Deep dive
1. DeepSeek turned frontier reasoning into open infrastructure
Marc’s starting point is deliberately two-sided: America remains “by far and away” the intellectual leader in AI, drawing on neural-network research extending back to the 1940s, but DeepSeek delivered an exceptionally good implementation and then performed the consequential act—“they gave it to the world in the form of open source.”
The inversion is in the names and behavior. OpenAI is now closed enough that Elon Musk’s lawsuit demands it be renamed “ClosedAI,” Anthropic has stopped publishing research, while DeepSeek released code and detailed technical papers for its V3 LLM and R1 reasoner, effectively publishing a roadmap for anyone attempting comparable work.
Patrick’s data-security concern gets an important distinction: using DeepSeek’s website may send data to China, but downloaded weights can run under the user’s control. Perplexity hosts R1 in the United States, Microsoft and Amazon offer it in American clouds, and the full model can run locally on roughly $6,000 of hardware.
Distillation extends that abundance down the stack. Smaller, somewhat less capable derivatives already run on MacBooks and iPhones and can be specialized for particular tasks, converting something “super esoteric six months ago and ultra-expensive and proprietary” into permanently accessible infrastructure for creative work, coding, mathematics, science, law, finance, and medicine.
2. A 30X cost collapse can hurt incumbents while enlarging their markets
On a static scorecard, every user wins. Marc’s example is an AI-lawyer startup whose core input became 30X cheaper in one week: like gasoline falling to one-thirtieth its price, the company can travel farther on the same budget, expand capability, or make the service cheap or free.
The obvious near-term losers are OpenAI, Anthropic, and other proprietary labs. Their defensive public messages looked “strong but clearly provoked,” prompting Marc’s invocation of the old rule: “When you’re explaining, you’re losing.” DeepSeek made whatever those companies offered the week before insufficient, which is creative destruction doing its work.
NVIDIA also faces a direct efficiency challenge. Its standard AI chips can carry margins as high as 90%, but DeepSeek documented how to achieve comparable work with fewer, cheaper NVIDIA chips; China’s emerging domestic chip supply presents the more fundamental threat to a company whose valuation reflects its dominant position.
The dynamic case may reverse every static conclusion. Through Jevons Paradox, a 30X cost reduction might elicit 30X, 100X, or 1,000X usage, just as new highways quickly fill with induced traffic. DeepSeek, proprietary labs, NVIDIA, and Chinese chipmakers could all grow because reasoning is “fire being brought down from the mountain” before most people have learned to use it.
3. China’s open-source judo move reframed Cold War 2.0
Marc’s Sputnik analogy reaches beyond a model benchmark. The first Cold War was military, economic, ideological, and technological; the Soviet inability to build a computer industry meant it was “ultimately doomed basically the minute the microprocessor was invented,” even though the consequence took roughly 20 years to resolve.
U.S.–China competition has those same dimensions but far deeper commercial interdependence. Huawei demonstrated the global-stack contest by offering telecom infrastructure at roughly one-third the Western price; exports generated revenue, while control of routers, switches, and wireless systems could also enable surveillance and geopolitical leverage.
AI is the next stack every country will integrate into healthcare, education, transportation, telecom, computing, and energy. That makes the choice between U.S.-derived and Chinese-derived systems more consequential than an ordinary vendor decision: each side wants its technology, political assumptions, and national champions embedded in the world’s operating fabric.
DeepSeek flipped the expected alignment. China took the lead in offering “the best open-source system,” while Western labs and policymakers favored black boxes, restrictions, and censorship. Marc calls it a generous act but also a “judo move”: unless the West returns to its own principles, China may “seize that mantle,” an outcome he considers ruinously destructive for Western power.
4. AI ownership claims collide with a long lineage and secret datasets
Patrick presses the apparent unfairness: Chinese developers can build abundance on discoveries financed by American capital. Marc answers that OpenAI did not invent the transformer; Google published it in 2017, left it unproductized for five years over safety concerns, and itself built on neural-network work dating to 1943.
Anthropic spun out of OpenAI, and every current lab inherited decades of university and government-funded work. Marc therefore rejects any “special moral claim” that today’s companies built AI de novo and deserve complete control; complaints about DeepSeek sound to him like understandable frustration after something already “out” and “done.”
The strongest alternative explanation remains explicitly unproven. DeepSeek’s surprising strength in English creative writing might reflect training on repositories such as LibGen, where pirated books are widely available, while U.S. labs may feel less able to use them—or DeepSeek may simply encode similar data more effectively.
Copyright law must decide whether training resembles copying a book or reading and learning from it. The practical asymmetry worries Marc more: if China trains on every book while American labs cannot, that “may be a death blow.” Nobody can resolve the comparison cleanly because DeepSeek releases weights, not training data, while OpenAI and Anthropic also keep both datasets and algorithms secret.
5. Maximum competition requires open source as a check on black boxes
Marc distinguishes the founder’s determinate optimism from the VC’s indeterminate version. A CEO must choose one plan and accept its trade-offs; a venture investor can fund 100 companies with mutually contradictory assumptions, leaving Marc philosophically free to favor “maximum competition, maximum freedom” and the fastest rate of evolution.
That means letting OpenAI and Anthropic run hard, funding startups of every shape, and protecting open source—provided large labs receive no preferential policy, subsidy, or government support. If open releases destroy particular business models, he accepts the trade: “The benefit to the world and to the industry as a whole is so big, we’ll find other ways to make money.”
Open source also prevents AI from becoming a black box controlled by a few companies aligned with government. Llama, Mistral, and now DeepSeek restore universities unable to purchase $1 billion of NVIDIA hardware to the field, allowing professors and incoming students to inspect code, teach it, research it, and publish the next breakthroughs.
6. Export controls can manufacture the rival they were meant to restrain
Leading-edge NVIDIA chips were already prohibited for sale to China, while a Biden executive order had proposed analogous software restrictions and, Marc thought, had since been revoked. He separates the legitimate national-interest argument from political implementation, where incumbents can weaponize government against open source or foreign competition: “When a rational argument encounters the political process,” rationality usually loses.
The first unintended consequence is forced efficiency. Denied the best chips, Chinese engineers optimized the legal, cheaper ones; that pressure became part of DeepSeek’s breakthrough and the reason a model of this class can run on $6,000 of hardware. The embargo may therefore have “already backfired in a significant way.”
The second is substitution. China now has overwhelming incentive to build a parallel chip industry that might take five years but eventually compete worldwide, sell below cost, and embed its own backdoors. Taiwan adds a more dangerous possibility: controlling its “magic fabs” could become an additional reason for China to invade rather than buy chips commercially.
Patrick’s sharpest challenge concerns Chinese software collecting American behavioral data. Marc answers with his “single-celled” objective—“I want America to win”—but defines victory as peaceful Cold War success through strength, openness, democracy, and trade. His preferred strategy is to “seduce the world into being on our side,” not sabotage America by becoming more controlling and threatening.
7. AI will standardize analysis before it replaces founder judgment
Marc expects reasoning models to transform late-stage and public-market work. Buffett’s reputed preference for annual reports over charismatic CEOs—the “great hair, great teeth,” polished shoes, and seductive story—creates an ideal machine task: o1, o3, o7, R4, or their successors will probably analyze regulated disclosures better than most manual investors.
That edge will not remain proprietary for long. Investment management is an arms race: once model-assisted analysis works for one firm, competitors adopt it, the arbitrage closes, and the capability becomes standard. Marc concedes that AI-focused venture firms could still become the shoemaker whose son has no shoes, so a16z has multiple internal efforts underway.
Early-stage venture may prove harder to automate because its first five years require face-to-face evaluation, coaching, and mutual learning. Marc compares this “project picking” to Queen Isabella judging Columbus: an old alchemy joins idiosyncratic people with enablers willing to fund uncertain journeys. Still, he hedges cleanly: “If I meet the algorithm that can do that better than I can, I will instantly retire.”
8. a16z pairs timeless founder service with an adaptive operating company
The timeless layer is respect for the entrepreneur: know the details, avoid “drive-bys,” build face-to-face relationships, and expect companies to take 10, 20, or 30 years rather than arrive overnight. NVIDIA, approaching what Marc thought was its 40th anniversary with an original VC possibly still on the board, exemplifies that duration.
Everything surrounding those values must move quickly. Venture’s old rule against funding PhD researchers failed once researchers founded leading AI companies; a16z went from no political function four years earlier to what Marc calls one of business’s largest and most sophisticated political operations. It is also asking what an org chart looks like when a company “actually fully uses AI.”
Marc contrasts a traditional six-GP partnership—assistants, associates, “eat what you kill,” difficult personnel decisions, and little franchise value—with a real company. The latter requires management, operating tempo, training, performance systems, specialization, and institutional assets that persist beyond whichever partners currently occupy the seats.
Scale is meant to support founders, not turn a16z into an asset gatherer harvesting fees. Growth funds can supply the hundreds of millions some AI companies require, but early-stage venture remains the core: a $5 million Series A may rival a $500 million later investment in return potential, while supplying the relationships and insight that make the larger platform distinctive.
9. Counter-elites—not direct mass revolt—are displacing incumbents
Drawing on The Machiavellians, Marc invokes the “Iron Law of Oligarchy”: 350 million people cannot directly organize, so every workable political system places a small organized elite over a large disorganized mass. Republics mediate this reality through representatives; the arrangement fails when elites become abusive enough to summon “pitchforks and torches.”
Change usually comes not from spontaneous mass organization but from a counter-elite claiming it can represent those masses better. Bernie Sanders and AOC challenged the Democratic establishment; Trump and MAGA displaced the Bush-era Republican establishment, against overlapping incumbents sometimes described as a “uniparty.”
The same pattern appears abroad: Jeremy Corbyn challenged Britain from the left, Nigel Farage’s Reform Party threatens the collapsed Tories, and Germany’s AfD rose under Alice Weidel as the CDU cooperated with it on something for what Marc described as the first time in roughly 50 years. Marc sees “a pitch battle basically between old elites and new elites.”
Trump’s claim was never that a billionaire in a golden penthouse literally resembled a rural farmer; it was that he was a counter-elite able to represent that farmer better. Marc maps the identical transition onto media, where Patrick, Joe Rogan, and others draw audiences away from network news, cable, newspapers, and prestige magazines despite furious attacks from incumbent institutions.
10. Institutional trust fell as elites stopped renewing themselves
Gallup’s trust-in-institutions series, Marc notes, peaked in the late 1960s or early 1970s and declined long before the internet; examples in his discussion include roughly 11% approval for institutions such as CNN or Harvard, single-digit trust in network news, and collapsing viewership, suggesting not merely criticism but broad withdrawal of belief.
One explanation is false consciousness: citizens were misled by misinformation, populists, and demagogues and will return once corrected. The rival explanation is that institutions became “rotten and dysfunctional and corrupt” after too much unchallenged power. Marc allows “some of each” but gives substantial weight to genuine nonperformance and self-dealing.
Healthy elites avoid decay through a “circulation of elites,” recruiting promising outsiders both to refresh themselves and neutralize future challengers. Marc spent roughly 25 years being invited to Davos, Aspen, New York dinners, and conversations with The New York Times—an “ego bath” available so long as he never deviated from approved opinions.
Marc casts himself and J.D. Vance as recruits who eventually rejected the bargain after concluding the institutions were authoritarian, self-interested, and dishonest. His closing companion framework is Joseph Henrich’s The WEIRDest People in the World, which he recommends for understanding how cultures—especially Western cultures—shape globalization, immigration, and political conflict.
11. Politics became technology’s fourth and most restrictive curve
Marc originally tracked three lines: technology’s upward squiggle and occasional step-change, society’s slower curve of readiness, and finance’s “EKG of a heart attack” cycling between panic and euphoria. Entrepreneurs and investors win by finding moments when the invention works, people want it, and capital can fund both construction and exit.
Politics became the fourth line during the preceding five years. In Marc’s account, government was overwhelmingly the largest recent constraint: crypto policy was devastating, fintech was damaged, social media was pressured into censorship, and AI faced proposed controls. “My main enemy is my own government,” he says, describing that position as both bizarre and politically clarifying.
He dates the cultural break to roughly 2013. The Clinton-Gore “deal” had allowed a pro-business, pro-technology Democrat to build wealth, receive favorable coverage, and become a philanthropist; then even private giving acquired the slur “philanthrocapitalism,” while social media went from an unalloyed good in 2012 to an alleged destroyer of democracy by 2016.
Marc describes the Biden years as an “endless barrage” of investigations, prosecutions, Wells notices, debanking, censorship, and sector-wide attacks, while hoping the new administration ends that posture. His desired equilibrium is bipartisan: Democrats return toward the center rather than intensify their internal push toward more radical economic, technology, and social policy.
12. Global supply-chain efficiency is inseparable from fragility
The canonical iPhone contains parts from roughly 40 countries before final assembly at Foxconn. China gets to count the full export value of the finished device even when its economic value added is only a single-digit percentage, making bilateral trade statistics a poor representation of the actual web of subcomponents, components, and specialized labor.
That specialization is why an iPhone costs $1,000 rather than “a billion dollars,” raised U.S. living standards, and helped lift billions abroad from poverty. Reversing globalization is therefore not equivalent to relocating one steel mill; it means unwinding a 40-country system whose pieces repeatedly cross borders before becoming a product.
Security exposes the system’s weak links. An autonomous U.S. military would depend on leading-edge Taiwanese chips vulnerable to Chinese control, while COVID test-and-trace capacity was reportedly constrained by something as prosaic as a plastic test tip made in one Chinese factory. Sophistication is unnecessary when a missing cheap component can halt the whole system.
Offshoring also hollowed out communities across the Midwest and South whose residents “can still vote,” turning economic efficiency into political instability. Marc expects continued interdependence and tariff bargaining rather than simple reshoring: connectivity creates more routes for adaptation, yet also more ways for one broken node—through war or pandemic—to break everything.
13. China is climbing from phones to drones to cars to robots
Marc’s four-rung ladder is “phones, drones, cars, and robots.” Shenzhen’s thousands of specialist electronics and mechanical companies first supported billions of phones, then drones—effectively flying phones with overlapping components—and created an ecosystem competitors cannot reproduce merely by designing a superior standalone product.
China consequently holds “ninety-nine point something percent” of the global drone market and supplies well over 90% of drones used by the American military. Every device might theoretically become a surveillance backdoor or kamikaze weapon, making commercial dependence inseparable from defense exposure.
Electric autonomous cars are the next rung because they resemble “a rolling laptop on wheels” more than traditional combustion vehicles. Tesla’s service bay looks like an operating room rather than an oil-stained workshop; Chinese manufacturers now exploit related supply chains to produce advanced cars at one-third or one-fourth the price of U.S. equivalents.
Robots complete the progression. Unitree’s robot dog starts at $1,500 versus roughly $50,000–$100,000 for Boston Dynamics’ version, while Chinese humanoids are approaching much lower prices too. A billion $10,000–$20,000 robots could build houses and provide care—or turn hostile during conflict. Marc expects the market to tip within a few years, followed by “giant geopolitical drama.”
14. Biotech is making reproduction and intelligence increasingly designable
Technology no longer remains obscure for long because 50 million or 100 million early adopters discuss each novelty online. Marc nevertheless identifies biotech—life extension, embryo selection, and reproductive technology—as the cluster whose implications remain underappreciated relative to its accelerating feasibility.
Embryos derived from stem cells might let infertile or older couples have biological children much later in life. External gestation remains further away, but Marc asks what happens to birth rates if people can continue having children into their 60s or choose a dozen children without conventional gestational limits.
Gene editing raises the sharper boundary. Scientists are identifying hundreds of genes correlated with IQ, making intelligence augmentation at least theoretically imaginable. Marc takes no position on permission or prohibition, but insists that these possibilities are “damned interesting” and could shape health and social organization for hundreds of years.
15. Autonomous war changes both battlefield economics and the threshold for conflict
American dynamism now spans defense, education, energy, and a16z’s first nuclear investment, but Marc is most focused on AI and autonomy changing military affairs. Ukraine’s tragedy has simultaneously become a “living laboratory,” where Ukrainian and Russian forces rapidly iterate drone-based attack and defense while planners worldwide study the results.
Past wars rewarded the side with the most men and material; future wars may reward the most money and technology because machines increasingly fight machines. That could spare soldiers, sailors, marines, and pilots, but Patrick and Marc preserve the darker implication: reducing human cost might make governments more willing to initiate conflict.
China has a major program applying AI to military systems, as do the United States and other countries. The race is therefore not speculative—it is already changing procurement, force composition, and the relationship between industrial capacity, software autonomy, and national power.
One senior planner told Marc that a backpack-sized suicide drone capable of crossing a hill and destroying a tank is the greatest defense innovation since the stirrup extended a mounted warrior’s lethality. His estimate: 20 well-trained soldiers with drones could hold off thousands or tens of thousands of conventional troops—a fundamental repricing of attack and defense still awaiting full comprehension.