Twelve months in AI time now feels like a century. Jensen Huang jokes about it, but his punchline has teeth: a year ago he said inference would grow not 100× or 1,000×, but a billion×—and this year he says he underestimated it. Why? Because inference stopped being “one shot.” Models learned to think—to plan, use tools, search, critique, and try again. In Jensen’s framing, we now have three scaling laws—pre‑training (learn), post‑training (practice), and inference (think). Add agents that run in parallel and you don’t just get faster answers; you get compound intelligence.
Zoom out. Ray Kurzweil’s “law of accelerating returns” says technology builds on itself, so progress stacks exponentially, not linearly. Thirty linear steps gets you to 30. Thirty exponential steps gets you to a billion. We are squarely at that part of the chessboard where each move doubles the grains of rice—and suddenly the emperor’s granary looks tiny.
Below is the lay of the land—where acceleration is fastest, why it’s happening, what it means for builders, and how to ride it without getting flung off the curve.
From chatbots to thinking systems
A year ago, most people saw LLMs as flashy autocomplete. Today, they’re systems of models coordinating in real time: one reasons, another searches, a third writes code, a fourth verifies and rewrites. Huang’s mantra—“think before you answer”—captures the shift. The result isn’t just better outputs; it’s longer‑horizon competence. Planning. Tool use. Memory. Self‑correction.
That’s also why usage curves kinked upward. Two exponentials began to stack:
- Adoption: more users, more use cases.
- Compute per task: reasoning chains, verification, and multi‑agent workflows cost more cycles per answer.
This is the key mental model for the next five years: systems that deliberate win.
Builder’s takeaway: Design for reasoning paths, not one‑shot prompts. Instrument and verify. Treat outputs like hypotheses, not endings.
Compute is the new energy (and revenue per watt is king)
The AI boom isn’t a software story alone; it’s a power story. The world is building AI “factories”—data centers measured in gigawatts. In Huang’s words, “NVIDIA revenue correlates to power.” Two truths now drive decisions:
- Everything moves to accelerated compute. General‑purpose CPU fleets are getting refreshed into GPU‑accelerated and domain‑specific stacks. First because they must (classical workloads run cheaper/faster on accelerators), then because they unlock new capabilities (agents, long‑context, video/gen‑AI).
- Perf per watt beats sticker price. In a power‑limited world, the question becomes: How many tokens (or tasks) per watt do I get? You could hand someone a rival chip for free and they’d still choose the system with higher tokens‑per‑watt if their power, land, and shell are the bottlenecks. That’s why NVIDIA keeps pushing full‑stack co‑design (chips + interconnect + compilers + libraries + orchestration) on an annual cadence. The target isn’t just speed—it’s lowering the cost of intelligence per joule.
Builder’s takeaway: Optimize for tokens per watt. Co‑locate with cheap electrons. Expect an annual hardware cadence; architect for swapability.
Higher ed hits the wall
Tuition has climbed nearly an order of magnitude since the early 1980s while median wages and other costs lagged far behind. Public sentiment flipped accordingly: fewer Americans now say college is “very important,” and an increasing share of long‑term unemployed are degree holders. Meanwhile, curricula ossify while the frontier moves weekly. One dean quipped they could build a nuclear reactor on campus faster than they can change the curriculum.
Credentialing is unbundling. Networks, portfolios, alternative pathways (bootcamps, apprenticeships, AI tutors) compete with the four‑year bundle. The winners won’t be the schools with the grandest lecture halls; they’ll be the ones that teach how to learn in a world where content is abundant and reasoning is scarce.
Builder’s takeaway: Hire for proofs of work. Fund the “AI‑native university” stack: adaptive tutors, agent‑graded projects, apprenticeship marketplaces, and verifiable, skills‑first credentials.
The robots leave the lab
Two years ago, humanoids were a meme. Today, they’re a capital magnet and an enterprise pilot away from going mainstream. One top startup in this space vaulted to a multi‑tens‑of‑billions valuation, partnered with a massive real‑estate owner, and secured access to hundreds of millions of square feet across homes, offices, and warehouses to gather the real‑world data humanoids need. OpenAI, which paused robotics in 2021 to focus on ChatGPT, has sprinted back—explicitly arguing that embodiment will matter for general intelligence.
Amazon’s move is even more cunning: AI glasses for drivers double as a training harness. Today, they augment workers. Tomorrow, the video and task traces become the curriculum for delivery bots. Humans, in effect, are training their successors—on company time.
Builder’s takeaway: If your product touches the physical world, assume you’ll have a robotic distribution channel—so start capturing the demonstrations now.
Work and wealth in an agentic economy
The workweek debate is heating up: tech CEOs speculate about a 3‑day norm as AI copilots handle much of the drudgery. Maybe. The more honest framing: output per human is going up fast; whether hours come down depends on competition and labor markets. In parallel, agents will coordinate across SaaS, APIs, and internal data to do real work—overnight and at scale. Enterprises won’t hire “another analyst”; they’ll spin up another team of agents.
On the wealth side, tokenization is finally leaving the whitepaper phase. Major venues are preparing to list tokenized securities, which means 24/7 trading, instant settlement, and simple fractionalization of assets. You’ll see crypto rails in very traditional places. Your portfolio may soon include a mix of public equities and tokenized slices of real‑world assets—and your agent might rebalance it while you sleep.
Builder’s takeaway: Productize “agent labor.” Price per task, not per seat. In fintech, design for tokenized rails and machine‑to‑machine finance.
Health in fast‑forward
Two waves broke at once:
- Continuous monitoring: Apple’s watch gained FDA clearance for a monthly signal that flags likely hypertension—the “silent killer” that affects over a billion people globally. This is preventive care with a vibration on your wrist. Combine it with CGM, rings, and smart earbuds and you get a life feed your doctor never had.
- AI‑designed medicines: DeepMind’s spinout and others are moving AI‑invented drug candidates into human trials. In plain English: we’re compressing “find a molecule” from years into months. Add protein design, lab automation, and in‑silico trials and you get medicine on exponential timelines.
Builder’s takeaway: The killer healthcare products pair real‑time sensing with closed‑loop interventions—and prove outcomes. For biotech, treat foundation models and wet labs as a single system.
Sovereign AI, geopolitics, and the American brand
Huang’s sovereign‑AI thesis is simple: energy grids, networks, and AI infrastructure are now state‑level priorities. Every country will use global models and build national stacks to encode local language, culture, law, and industrial know‑how. The American advantage—if we keep it—is talent, openness, and the willingness to export the stack.
There’s one policy point worth saying out loud. America’s superpower is still the American Dream. If we make it harder for the world’s best to study, work, and stay, we’ll handicap the very system that created this boom. Staple green cards to STEM diplomas. Make it easy for founders and their families to plant roots here. Winning the AI race is mostly about attracting and compounding human capital.
Builder’s takeaway: Assume every major region will have a sovereign stack. Build products that bridge them—privacy‑preserving, policy‑aware, and deployable on multiple clouds.
What the next five years actually feel like
Jensen’s forward look is refreshingly concrete:
- We’ll fuse AI with robots. Expect useful, if imperfect, humanoids in homes, hospitals, and logistics.
- Each of us will have a persistent AI—our personal R2‑D2—running in the cloud, embodied in our car, phone, and home devices.
- Data processing—the bulk of enterprise compute—shifts from CPU‑centric SQL to accelerated (and increasingly AI‑assisted) pipelines.
- Digital twins move from factories to people—predictive health models that help you intervene before your body complains.
The practical advice? Get on the train. Don’t wait for perfect clarity at some future intersection; compounding systems reward those who board early and learn while speed increases.
The investor’s playbook (2025–2030)
- Bet on perf/watt. In power‑limited environments, tokens per watt is revenue.
- Back full‑stack thinkers. Teams that co‑design model ↔ compiler ↔ silicon ↔ interconnect ↔ orchestration will bend unit economics.
- Fund data + distillation moats. Specialized datasets, verifiers, and task‑specific distillation that cut inference cost by orders of magnitude.
- Design for agents. Long‑horizon planning, tool use, verification. Price per outcome. SLA the task, not the API call.
- Co‑locate with cheap energy. Compute follows electrons. Prefer sites with abundant power and fast permitting.
- Exploit the annual cadence. Assume yearly hardware leaps. Bake swap‑readiness into your stack.
- Instrument everything. Telemetry for reasoning steps, failures, and human‑in‑the‑loop corrections. Verification is the moat.
- Tokenize responsibly. Structure compliant, 24/7 products where agents can operate.
- Talent arbitrage. Hire globally; lobby locally. Immigration is strategy.
- Ship now. In accelerating markets, time‑to‑learning beats time‑to‑perfection.
One more Kurzweil zoom‑out
Kurzweil once quipped that the 21st century won’t deliver “100 years” of progress but something like 20,000 years at today’s rate. That’s not prophecy; it’s a reminder that compounding is a harsh teacher. Our job—founders, investors, policymakers—is to bend the curve toward broad prosperity. Or, to channel Jensen’s bluntest advice: when the train starts to go exponential, get on it.
Sources
- WTF Just Happened in Tech slides (9/23 edition). Key visuals referenced: tuition vs. other costs; AI data‑center capacity outlook; model‑benchmark one‑upmanship; robotics TAM; employment and degree perceptions.
- NVIDIA: OpenAI, Future of Compute, and the American Dream | BG2 w/ Bill Gurley & Brad Gerstner (transcript). Anecdotes and framing used throughout: three scaling laws; “think before you answer”; two exponentials (adoption × compute per task); tokens‑per‑watt; annual release cadence (Hopper → Blackwell → Vera Rubin → Ultra → Feynman); AI factories; sovereign AI; personal R2‑D2; advice to “get on the train.”
- Ray Kurzweil: The Law of Accelerating Returns; The Singularity Is Near; The Singularity Is Nearer. Frameworks: exponential growth, accelerating returns, human‑level AI timelines, tech convergence.
- Public announcements from Apple (hypertension alert on Apple Watch) and DeepMind/Isomorphic Labs (AI‑designed drug programs) illustrating acceleration in healthcare and drug discovery.