Reid Hoffman on DeepSeek, AI investment, and political risk

Executive overview

The race for AI dominance is more competitive than most investors assumed, and DeepSeek exposed that. The cheap-model narrative is overstated — large-scale compute still underpins small-model training — but the signal that AI cost structures will compress is real.

Reid Hoffman argues AI is a net positive for humanity at the scale of electricity or the internet, not a zero-sum tech race. Political risk — retribution, chilling effects on dissent, federal workforce disruption — is the undercurrent throughout.

The biggest near-term AI risk isn't capability; it's who controls the institutions shaping its application.

DeepSeek: what actually happened

• DeepSeek's cheap-training claim is almost certainly incomplete — large models likely assisted training of the smaller one
• Training small models efficiently still requires large models; this doesn't reduce the need for scale compute
• The "$5–6M training run" figure excludes talent and infrastructure; no serious analyst believes it
• Market panic over Nvidia was a short-termism overreaction — AI capex payoff is 3–7 years, not quarters
• The real signal: capable AI models will get cheaper, compressing OPEX for every sector using AI

AI market structure

• More likely outcome is 7–15 major tech players, not winner-takes-all, closer to internet dynamics than PC era
• Network effects and enterprise integration will still favour incumbents, but a wide startup field will coexist
• Agent universe will be more agents than people — specialised by use case, personality, and user type
• "Super agency": when millions of people gain cognitive superpowers simultaneously, collective capability compounds
• AI as a platform mirrors electricity — it will power applications we cannot yet predict

Federal workforce and political risk

• The "fork in the road" buyout mirrors Elon's Twitter playbook — broad exits select out the best talent
• In private companies, retention incentives balance exit offers; no equivalent exists in the federal system
• Blunt-force workforce cuts can be justified in principle, but performance-based targeting is the right method
• Settlements like Meta's $25M to Trump set a chilling precedent — practical for shareholders, corrosive for institutions
• Removing security details from former officials who acted in service of the country is "repackaged violence"

RFK Jr. and AI in health

• Reid's assessment: RFK confirmation will be "measured in thousands of American lives lost"
• RFK is anti-science in practice, not just anti-vaccine — ignores expert input in direct meetings
• HPV vaccine (Gardasil) reduces cervical cancer by 90%; HHS leadership that undermines it causes direct measurable harm
• Senators voting for confirmation should track the human cost of that calculus
• AI cancer research (Reid's Menace investment) treats cancer as a language misfiring — AI decodes the biology

Predictions

• AI agents will trend toward high-EQ, kind interactions — mental health benefits will emerge from market dynamics, not regulation
• Companies that budgeted $100–500M in AI OPEX will see 30–80% cost reductions from cheaper models — releasing reserves back to earnings
• Podcasters and independent journalists entering the White House briefing room is a legitimate broadening of access