SpaceX signs $6.3B compute lease with Reflection for GB300 access. Reflection has signed a major compute agreement with SpaceXAI, giving the Nvidia-backed open-source AI startup access to GB300 chips and other hardware inside the Colossus 2 data center. Under the compute agreement reported by Axios, Reflection will pay SpaceXAI $150 million per month starting July 1, 2026, after an initial ramp period. The deal runs through 2029 and is worth about $6.3 billion if maintained through the full term. Either party can terminate the agreement with 90 days' notice after the first three months. That structure gives Reflection access to high-end infrastructure without building a full hyperscale data center itself, while giving SpaceXAI another large customer for its rapidly expanding compute business. The deal matters because open-source AI developers are trying to compete in a market where compute access has become one of the hardest constraints. Reflection is still training its models, and Colossus 2 access gives the company a larger hardware base as frontier model development increasingly depends on full clusters, power, cooling and deployment speed rather than chips alone. SpaceXAI adds another compute customer. Reflection joins a growing list of AI customers tied to Elon Musk's compute infrastructure. Anthropic previously tapped SpaceX's Colossus 1 for large-scale Claude capacity, while Google signed a separate agreement to pay SpaceX $920 million per month for cloud compute. Reuters reported that Google's agreement gives it access to about 110,000 Nvidia GPUs, CPUs, memory and related compute infrastructure from October 2026 through June 2029. SpaceX's compute agreements with Anthropic and Google were already valued at more than $70 billion in aggregate if held to term, before the Reflection deal added another multibillion-dollar customer. The Reflection agreement also fits the newer customer mix. Anthropic needs capacity for Claude. Google wants large-scale AI infrastructure. Cursor entered SpaceX's orbit through a separate AI coding push after the Anysphere acquisition. Reflection brings the open-source model angle, putting SpaceXAI capacity behind a company trying to compete with closed frontier labs. Compute becomes the AI market bottleneck. The deal reinforces how AI competition is moving from model announcements into infrastructure access. Nvidia chips remain scarce, but the bigger constraint is often complete deployed capacity: data-center space, power, networking, cooling, cluster reliability and the financial ability to reserve hardware for years. Reflection's arrangement also shows how Nvidia's ecosystem is feeding its next wave of customers. Axios reported that Nvidia invested $800 million in Reflection, which is now leasing access to Nvidia chips purchased by SpaceXAI. That circular structure is becoming more common across the AI market, where chipmakers, cloud operators, model developers and infrastructure owners are increasingly suppliers, investors and customers to one another. SpaceX's AI infrastructure expansion has already drawn capital-market scrutiny. The company has been linked to a possible post-IPO bond sale as investors weigh how much debt and compute revenue can support its AI buildout beside launch, Starlink and satellite infrastructure. The Reflection lease adds another recurring revenue stream to that model. The agreement gives Reflection GB300 capacity at Colossus 2, gives SpaceXAI another long-duration compute customer, and pushes the AI race further toward companies that can secure power, chips and deployed clusters before the next model cycle begins.

Pentagon AI contracts skip Anthropic and Nvidia loses China. Robert Hattala May 5, 2026 p>Big day in AI news, and not the kind where someone ships a new model with a clever logo. This was the kind of day where the bills come due and the politics show up. Three stories worth your time today. Pentagon money, Nvidia's China problem, and a real medical win from Mayo Clinic. Let's get into it. Pentagon hands out AI contracts and Anthropic walks. On May 1, the Pentagon awarded classified-network AI contracts to AWS, Google, Microsoft, Nvidia, OpenAI, SpaceX, and Reflection AI. Anthropic was not on the list. The why is the part that matters. Anthropic refused to let Claude be used for "all lawful" purposes. They flagged the language as too broad, saying it could open the door to domestic mass surveillance or autonomous weapons. Why it matters: there's a real split now between AI labs willing to take any defense dollar and labs that draw a line in the sand. The Pentagon wrote a blank check, and one company said no thanks. My take: I respect Anthropic for this. You can't build a "safe AI" brand and then hand the keys to whatever a contracting officer feels like doing on a Tuesday. The dollars are big. The trust they keep is bigger. Other labs ought to pay attention because customers in regulated industries are watching too. Nvidia's China AI share just hit zero. Jensen Huang said this week that Nvidia's market share for AI accelerators in China is now zero. Huawei, Cambricon, and other domestic chip outfits are filling the gap. Huang argued the U.S. export controls have "already largely backfired." Hard to disagree when your share goes from dominant to nothing in a couple of years. Why it matters: export controls were supposed to slow China down. Instead they pushed China to build its own stack. Now there are two separate AI hardware worlds, and the American one no longer ships into the bigger market. My take: this was always the risk with broad export controls. You don't strangle a country with this much capital and this many engineers. You just give them a five-year head start on import substitution. The U.S. lost a customer and gained a competitor in the same trade. That's a bad day on the chess board. Mayo Clinic says AI can spot pancreatic cancer years early. Mayo Clinic researchers showed an AI model that flags pancreatic cancer years before clinical diagnosis. Pancreatic is one of the worst because by the time you feel something, it's usually late. Why it matters: this is the kind of work that justifies all the noise around AI in medicine. Not chatbots that summarize charts. Not slick demos. Pattern recognition on imaging and labs that actually saves lives. My take: more of this please. Less PR theater about agents that can book your dentist appointment, more boring research that catches a cancer two years before a human radiologist would. Mayo is doing it right by publishing real results instead of dropping a press release and hiding the data. Putting it together. One day, three signals. AI vendors are getting sorted into camps based on who they will sell to. Hardware policy is shifting under everybody's feet. And the actual humans-helping-humans use cases are quietly winning while the rest of the field argues on Twitter. Worth paying attention to all of it.

Shinsegae Group and Reflection AI to build 250MW sovereign AI data center in South Korea. Will offer compute to government and enterprise clients March 19, 2026 Shinsegae Group and Reflection AI are teaming up to build a 250MW data center in South Korea, which they claim will be the country's largest sovereign AI factory. Korean conglomerate Shinsegae and AI lab Reflection have signed a Memorandum of Understanding (MoU) laying out their intention to set up a joint venture (JV) to build the data center. The MoU was signed in San Francisco earlier this week at a ceremony attended by US Commerce Secretary Howard Lutnick. A Shinsegae department store in Centum City, Busan, Korea - Modamoda/Wikimedia Commons Under the JV agreement, Shinsegae will secure a plot of land and build the data center, for which Reflection AI will be responsible for the design and operation of the equipment inside. The project is expected to cost at least 10 trillion won ($6.8bn). Further details haven't been shared. "The Reflection and Shinsegae Group AI factory will equip the Korean government and enterprises with fully sovereign frontier capabilities built and operated on home soil, setting a new standard in technological self-sufficiency," a statement from the companies said. Shinsegae Group operates department stores, as well as e-commerce, hospitality, and real estate businesses, and is apparently keen to become South Korea's answer to Amazon and the other hyperscalers. "The data center that we are building in partnership with Reflection AI will not only be a growth opportunity for Shinsegae, but also a pivot point for the Korean AI ecosystem," said Yongjin Chun, chairman of Shinsegae Group. Reflection AI is developing open AI models akin to China's DeepSeek, rather than the proprietary systems owned by the likes of OpenAI and Anthropic. It was set up by Misha Laskin and Ioannis Antonoglou, two former Google DeepMind engineers. It is backed by Nvidia and raised $2 billion in October 2025. Earlier this month, it was reported to be seeking further funding in a round that would value the business at $20bn. In November, it announced a deal to use GMI Cloud's US-based infrastructure to train and run its models. It is also one of a number of companies that are part of an Nvidia-founded open AI coalition, Nemotron, which was launched at the GPU giant's GTC event in California earlier this week. Get a roundup of the latest regional news across asia fortnightly. More in construction & site selection.