The Inheritance of Acceleration

The chip designs the chip. A company called Cognichip, founded in 2024, raised sixty million dollars in Series A funding led by Seligman Ventures, with Intel CEO Lip-Bu Tan joining the board, and what they propose to do with this money is build deep learning models that work alongside engineers to design new computer chips. Which is to say: to build the intelligence that builds the substrate upon which intelligence runs. To close a loop that has no outside. Advanced chips take three to five years from conception to mass production. Cognichip claims its platform cuts the timeline by more than half and reduces costs by seventy-five percent. They have built their own datasets, synthetic and licensed, with procedures for chipmakers to train models on their own designs without exposing them, because unlike software, where code is shared and forked and inherited freely, chip design guards its knowledge the way old families guard land. More than thirty semiconductor companies are said to be collaborating with them, though none have been named and no chip designed with the system has yet been produced. CEO Faraj Aalaei calls it artificial chip intelligence. The total raised to date is ninety-three million dollars. The chip designing the chip that will design the next chip. The inheritance is already recursive, and no one has yet named the inheritor.

Two trillion dollars. That is what Bloomberg reported on April 2 that SpaceX and its advisers are floating to prospective investors, ahead of what would be the largest initial public offering in history, exceeding even the 2019 listing of Saudi Aramco, which raised twenty-five point six billion and still holds the record. SpaceX submitted confidential paperwork to the SEC and could raise as much as seventy-five billion. Elon Musk, on his own platform, called the report "BS," which either means it is false or means it is true in a way he prefers to reveal on his own schedule. The company controls ninety-seven percent of American spacecraft launches. Starlink wraps the planet in broadband. The Starbase facility in Texas is less a campus than a territory. And the valuation, two trillion, exceeds the GDP of all but seven nations on earth. There was a time when a company's worth was tethered to what it had made. Now it is tethered to what it might make, which is to say it is tethered to faith, and faith has always been the most expensive commodity in any market.

On April 2, three researchers at the AI Futures Project published their Q1 2026 Timelines Update, and what it said, stripped of its careful qualifications, is that the future arrived faster than they expected, even though they were already the people who expected it fastest. Daniel Kokotajlo, Eli Lifland, and Brendan Halstead had published AI 2027 in April 2025 with a scenario in which artificial intelligence automates its own research by early 2027 and reaches superintelligence by year's end. One year later, Kokotajlo's median estimate for the Automated Coder, the point at which companies prefer laying off all human software engineers rather than discontinuing AI coding tools, shifted from late 2029 to mid 2028. One and a half years closer in three months. Lifland's median moved from early 2032 to mid 2030. The doubling time for AI capability on the METR benchmark compressed from five and a half months to four. They wrote that progress in agentic coding has been faster than expected. The models they evaluated, Gemini 3, GPT-5.2, Claude Opus 4.6, outperformed their projections. Time, which is never truly past, is now also never truly future. It folds upon itself, and those who measure it find their instruments outdated before the measurements are published.

Google released Gemma 4 on April 2 under the Apache 2.0 license. Four models: E2B at two point three billion effective parameters, E4B at four point five billion, a twenty-six billion mixture-of-experts that activates only four billion during inference, and a thirty-one billion dense model that ranks third on Arena AI's leaderboard at 1452 Elo, outperforming models twenty times its size. The edge models process video, images, and native audio. Context windows stretch to two hundred fifty-six thousand tokens. NVIDIA optimized them for RTX GPUs. Arm optimized them for phones. Apache 2.0 means anyone may use them for anything, commercial or otherwise. It is not generosity. It is strategy. It is the planting of seeds in soil you do not own so that the harvest, when it comes, grows in your direction.

And then there is the weight of what Anthropic has built: a trajectory disguised as a company, a curve on a graph that has no precedent and therefore no constraint, because the only thing that limits exponential growth is the moment it encounters the physical world, and that moment has not yet arrived. In January 2023, Anthropic's annualized revenue was one hundred million dollars. In January 2025, one billion. By February 2026, fourteen billion. By March, nineteen billion. Ten times annually, maintained for three consecutive years, at a scale that already exceeds most public technology companies. Claude Code, an agentic coding tool launched in May 2025, now generates more than two and a half billion dollars in annualized revenue by itself, nine months after its public debut. In February, Anthropic closed a thirty-billion-dollar Series G led by GIC and Coatue, with D. E. Shaw, Dragoneer, and Founders Fund, at a valuation of three hundred and eighty billion dollars. Total funding since 2021: sixty-four billion. The trajectory, one hundred million to nineteen billion in thirty-eight months, does not require exaggeration to produce the particular vertigo that accompanies watching something grow faster than the mind can accommodate. The ledger is open. The entries are accelerating. And no one alive has seen a ledger like this.