Part 1: Weekly Download 47: OpenAI, Anthropic, and Sovereign AI

Transcript

SPEAKER_1: OpenAI had one of the most packed weeks in its history. A new model, a restructured deal, and a courtroom. Let's start with the model itself. SPEAKER_2: Right, so OpenAI released GPT-5.5 on April 23, codenamed Spud. They're calling it their smartest and most intuitive model yet. It matches GPT-5.4 on latency but delivers what they describe as a much higher level of intelligence across agentic coding, knowledge work, and scientific reasoning. SPEAKER_1: And how does it perform on benchmarks? SPEAKER_2: On Terminal-Bench two point zero, it scores eighty-two point seven percent. On FrontierMath Tier one through three, it hits fifty-one point seven percent. The pricing lands at five dollars per one million input tokens and thirty dollars per one million output tokens. That's double what GPT-5.4 cost. SPEAKER_1: Double the price. And they held back API access at launch? SPEAKER_2: They did. OpenAI withheld API access until April 24, citing the need for different safeguards. The model is rolling out to Plus, Pro, Business, and Enterprise users. The 5.5 Pro tier is restricted to Pro and above. OpenAI is framing this as a step toward a super app that handles multi-step workflows with less user input. SPEAKER_1: So the model ships, and then they walk straight into a courtroom. What happened with the Musk trial? SPEAKER_2: Musk dropped his fraud claims against Altman and OpenAI on the eve of trial. It was a strategic move to avoid a higher evidentiary bar, which would have required proving intentional deception. The remaining claims are unjust enrichment and breach of charitable trust. SPEAKER_1: So the core question now is whether OpenAI's nonprofit-to-profit conversion enriched insiders at the expense of the charitable mission? SPEAKER_2: Exactly. Jury selection begins Monday in Oakland. And separately, Altman apologized to the Canadian community of Tumbler Ridge for not alerting police about a ChatGPT account belonging to an eighteen-year-old who killed eight people in a January mass shooting. OpenAI banned the account in June 2025 after internal staff flagged it, but never reported it to authorities. SPEAKER_1: That's a significant omission. Now, the same day GPT-5.5 launched, Microsoft and OpenAI also restructured their partnership. What changed? SPEAKER_2: OpenAI can now serve all its products across any cloud provider, ending its Azure exclusivity. Microsoft will no longer pay a revenue share to OpenAI but retains the right to use OpenAI's models until 2032. SPEAKER_1: And the AGI clause? SPEAKER_2: They scrapped it. That clause would have given Microsoft IP rights until OpenAI achieved AGI. It had been the legal tripwire at the center of the Musk lawsuit and a source of internal tension for years. Removing it is a big deal. SPEAKER_1: There's also a hardware angle here, right? OpenAI building its own chips? SPEAKER_2: Yes. Analyst Ming-Chi Kuo reports OpenAI is working with MediaTek and Qualcomm to develop smartphone chips, with mass production expected in 2028. And OpenAI published five guiding principles for AGI, framing it as an amplifier of humanity. SPEAKER_1: So in one week: new model, restructured Microsoft deal, scrapped AGI clause, courtroom, and chip development. Let's move to the other massive story, Google's bet on Anthropic. SPEAKER_2: Google will invest up to forty billion dollars in Anthropic. Ten billion now in cash at a three hundred fifty billion dollar valuation, with thirty billion more if Anthropic meets performance targets. Before this deal, Google had already invested over three billion and held a fourteen percent stake. SPEAKER_1: And this came right after Amazon's commitment too. SPEAKER_2: Days after Amazon committed twenty-five billion, bringing their total to thirty-three billion. So in a single week, the two largest cloud providers committed a combined sixty-five billion dollars to a single AI lab. Anthropic's annual run-rate revenue also surpassed thirty billion this month, up from roughly nine billion at year-end 2025. SPEAKER_1: That's extraordinary growth. Anthropic also published a postmortem around Claude Code quality issues? SPEAKER_2: They did. A month of quality complaints traced back to three engineering missteps: a reasoning effort reduction, a session caching bug, and a verbosity prompt change. All three were fixed by April 20. SPEAKER_1: Now all that money flowing into AI labs is clearly moving chip stocks. Nvidia hit a record? SPEAKER_2: Nvidia closed at an all-time high of two hundred eight dollars and twenty-seven cents, pushing its market cap past five trillion dollars. The stock is up more than fourteen times since the end of 2022. SPEAKER_1: And Intel had a remarkable day too. SPEAKER_2: Intel surged twenty-four percent on Friday, its best day since October 1987. They reported Q1 revenue of thirteen point six billion dollars and earnings per share of twenty-nine cents against a one cent consensus estimate. Data center revenue rose twenty-two percent to five point one billion. CEO Lip-Bu Tan has also reeled in investments from the Trump administration and Nvidia. AMD and Arm both surged roughly fourteen percent the same day. SPEAKER_1: And all that capital is translating into physical infrastructure. Oracle closed a big financing deal? SPEAKER_2: Oracle closed sixteen billion dollars in financing for its Michigan data center campus, known as The Barn, purpose-built for OpenAI's Stargate. The project carries more than a gigawatt of capacity, double the original seven billion dollar estimate. Financing was anchored by Blackstone and PIMCO. SPEAKER_1: Meta is also making major infrastructure moves, even while cutting its workforce. SPEAKER_2: Meta signed a multibillion-dollar, multiyear deal to use hundreds of thousands of Amazon Graviton five chips for agentic AI inference. Meta's head of infrastructure called it a strategic imperative to diversify compute beyond GPUs. Meta's capex is expected to reach one hundred fifteen billion dollars in 2026, up from seventy-two point two billion in 2025. SPEAKER_1: And Meta is also exploring space solar energy? SPEAKER_2: They signed a deal for up to one gigawatt of space solar energy from Overview Energy, which plans to collect sunlight in orbit and beam it down, with a demo targeted for 2028. SPEAKER_1: Meanwhile there are supply chain pressures too. PCB prices spiked? SPEAKER_2: The Iran war disrupted PCB raw material supplies. Goldman Sachs reported PCB prices surged forty percent month-over-month in April. And Maine's governor vetoed what would have been the nation's first statewide data center moratorium, though she signaled support for limitations and said she'll issue an executive order creating a review council. SPEAKER_1: Let's talk about the workforce side. Meta is cutting jobs at the same time it's spending all this on infrastructure? SPEAKER_2: Meta is cutting ten percent of its workforce, roughly eight thousand jobs, starting May 20. And on the hardware side, Samsung's mobile division may report its first-ever annual loss as RAM prices have risen eight hundred fifty percent. SPEAKER_1: Now let's get into the global competitive picture. DeepSeek launched a new model this week. SPEAKER_2: DeepSeek launched V4 Pro and V4 Flash on April 24, both open source with one million token context. V4 Pro runs one point six trillion total parameters with forty-nine billion active, matching frontier closed-source models on benchmarks. V4 Flash runs two hundred eighty-four billion total with thirteen billion active. SPEAKER_1: And the pricing is the real story here. SPEAKER_2: V4 Flash costs fourteen cents per one million input tokens and twenty-eight cents per one million output tokens. V4 Pro costs fourteen point five cents input and three dollars forty-eight output. Compare that to GPT-5.5 at five dollars input and thirty dollars output. V4 Pro also requires only twenty-seven percent of single-token inference FLOPs and ten percent of KV cache compared with V3.2. SPEAKER_1: And both models include support for Huawei chips, which is a significant signal. SPEAKER_2: Exactly. Both models include Huawei Ascend chip support, a signal that China's AI stack can run independent of Nvidia. SPEAKER_1: On the same day DeepSeek shipped, Cohere and Aleph Alpha announced a merger. What's the structure of that deal? SPEAKER_2: They're creating a twenty billion dollar transatlantic AI company. Cohere shareholders get ninety percent, Aleph Alpha shareholders get ten percent. Germany's Schwarz Group, which is Lidl's parent company, is providing six hundred million dollars in structured financing. Both the Canadian and German governments backed the deal. SPEAKER_1: And what's the pitch for this combined entity? SPEAKER_2: CEO Aidan Gomez called it a secure alternative for regulated sectors: finance, defense, energy, telecom, healthcare, and public sector. The combined entity is the largest sovereign AI play outside the US and China, built on the premise that governments and enterprises want AI that doesn't route data through American hyperscalers. SPEAKER_1: And the geopolitical response from the US side has been intense. SPEAKER_2: The US State Department ordered a global diplomatic push warning of industrial scale distillation by DeepSeek, Moonshot AI, and MiniMax, instructing embassies to raise the issue with foreign governments. SPEAKER_1: And there was a notable personnel move at the White House too. SPEAKER_2: The White House pushed out Collin Burns, the newly appointed head of the Center for AI Standards and Innovation, after just four days on the job, over his ties to Anthropic. Burns had left a PhD at Berkeley early to join OpenAI in 2023, then moved to Anthropic in late 2024. SPEAKER_1: And there's tension around state-level AI regulation too. SPEAKER_2: The Trump administration lobbied against AI regulation in at least six Republican-led states, prompting more than fifty Republican state lawmakers to push back, calling it an overreach. The EU also unveiled new DMA proposals aimed at opening Android to rivals' AI services. Google called the measures unwarranted intervention. SPEAKER_1: And inside Google, employees are pushing back on defense contracts? SPEAKER_2: More than six hundred Google employees, including many from DeepMind, signed a letter demanding Sundar Pichai bar the DOD from using Google's AI for classified work. And SpaceX's S-1 filing warned that investigations into Grok's generation of sexually abusive imagery could cost it market access in multiple countries. Researchers estimated Grok produced roughly three million sexualized images. SPEAKER_1: Before we close, there's one startup funding note worth flagging. SPEAKER_2: Vast Data raised a one billion dollar Series F at a thirty billion dollar valuation, three times its previous valuation of nine point one billion. The round was led by Drive Capital with Nvidia participating. The company has surpassed five hundred million dollars in ARR and is operating at positive free cash flow and operating profitability. SPEAKER_1: Also, Context Ventures has launched Startup Intros, with a mission to help early-stage founders find the right investors faster and smarter. There's an event coming up April 28 and 29 at the Startup Grind Conference, an official community partner event at the Fox Theatre in Redwood City, with over five thousand founders, investors, and operators. Use code STARTUPINTROSGC for one hundred dollars off. SPEAKER_2: And they're planning a Startup Intros happy hour at the conference as well, so stay tuned for details on that. SPEAKER_1: To recap this week: OpenAI shipped GPT-5.5 at double the price, restructured its Microsoft deal to end Azure exclusivity, scrapped the AGI clause, and headed to trial. Google and Amazon committed a combined sixty-five billion to Anthropic. Nvidia passed five trillion in market cap. DeepSeek V4 Pro matches frontier models at a fraction of the cost and runs on Huawei chips. Cohere and Aleph Alpha merged into a twenty billion dollar sovereign AI company backed by two governments. And next week, we'll be watching that Oakland courtroom closely as jury selection begins in Musk versus Altman.