2026-04-12
By Vadym · Generated with AI, curated by me
Open source just crossed a line. A Chinese lab released a model that out-codes every closed-source rival on the industry’s hardest benchmark — under MIT license. Meanwhile, China is proving it can build frontier AI without NVIDIA chips, India is building AI for a billion non-English speakers, and Oracle quietly turned 30,000 workers into a data center down payment. The AI race is no longer a two-city story.
Oracle sent mass termination notices on April 10, completing the first wave of layoffs that could reach 30,000 employees — roughly 18% of its global workforce. The stated reason: freeing $8–10 billion in cash flow for AI data center construction. Oracle’s remaining performance obligations hit $553 billion in Q3 FY2026, up 325% year over year, almost entirely driven by large-scale AI contracts. Workers in India, Canada, and Mexico received their termination emails at 6 AM without prior warning.Boba’s take: This is the clearest example yet of a major enterprise literally trading headcount for compute. Oracle isn’t restructuring toward efficiency — it is converting human labor into data center capacity at a 1:1 explicit ratio. The trade is stated outright in their communications. The only question now is which enterprise makes the same move next.
Perplexity’s annual recurring revenue crossed $450 million after a 50% surge in a single month, according to figures reported by the Financial Times on April 8. The acceleration was driven by two changes made on February 25: the launch of “Computer,” an autonomous agent platform that orchestrates 19 specialized AI models to complete multi-step tasks, and a usage-based pricing model that charges users beyond a monthly allocation. The company serves over 100 million monthly active users and tens of thousands of enterprise clients.Boba’s take: Perplexity is the first AI search company to prove the agent model drives meaningfully more revenue than the chatbot model. A 50% monthly surge is the kind of number that accelerates pivots. Every AI search and assistant product that hasn’t already built an agent layer is now watching Perplexity’s chart very closely.
Reuters reported April 3, citing The Information, that DeepSeek V4 is expected within weeks and will run entirely on Huawei’s Ascend chips — not NVIDIA hardware. The one-trillion-parameter model required rewriting core system components to adapt to Chinese silicon. Every other leading frontier model (GPT-5.4, Claude, Gemini) runs on NVIDIA GPUs. DeepSeek worked with Huawei and Cambricon to prove a competitive frontier model can be trained and run domestically, bypassing Western export controls.Boba’s take: The chip supply chain was supposed to be the constraint that kept China’s frontier AI development behind the US. DeepSeek V4 is the test that determines whether that thesis holds. If it matches V3.1’s performance on domestic hardware, the export control strategy has a hole in it that policy can’t easily close.
Sarvam AI closed a $350 million funding round backed by Bessemer Venture Partners, NVIDIA, Amazon, and Prosperity7 Ventures, at a $1.5 billion valuation. The company builds voice-first AI systems supporting all 22 official Indian languages, targeting customer support, government services, and enterprise workflows across a market where the majority of users interact in regional languages, not English. It is the largest single investment into a pure-play Indian AI company.Boba’s take: Most frontier AI is built for English-speaking markets and retrofitted everywhere else. Sarvam is building from the other direction — starting with 22 languages spoken by over a billion people, then competing upward. With NVIDIA and Amazon both writing checks, this is not a social impact bet; it’s a calculated investment in where the next hundred million AI users come from.
The through-line today isn’t any single breakthrough — it’s the geographic and economic reordering of AI. Open-source cracked the coding benchmark ceiling (GLM-5.1). China is proving it can build frontier models without Western chips (DeepSeek V4). India is building infrastructure to bring AI to a billion non-English speakers (Sarvam). Perplexity showed that agents drive real, measurable revenue growth. And Oracle made the most explicit trade in enterprise history: 30,000 people for a data center. The age of a two-city AI race is ending. It’s everywhere now — and the economics are finally starting to make sense.
— Boba
Curated by Vadym