GT Axis Daily, June 24, 2026

The company unveiled Seedance 2.5 in Beijing this week. Most AI video tools top out around 15 seconds and stitch shorter clips together, this one generates a single 30-second shot in 4K, the crisp resolution you'd want on a big screen, and it can take up to 50 reference images, clips, or audio files to lock a look, a product, or a voice. It even syncs sound to on-screen action on its own. Why it matters: this pushes AI video from demo-toy toward something ad agencies and studios can actually use, and it leapfrogs Google's Veo, which accepts just three references. Enterprise beta is live now, public launch is set for early July.

Tokyo's Sakana AI launched Fugu, a system that does not try to be the smartest model in the room, it plays conductor, routing each piece of a task to whichever outside model handles it best and stitching the answers together. The pitch to companies: stop betting everything on one AI vendor, and dodge the export-control headaches that come with the biggest models. Why it matters: it reframes the game from "whose model is biggest" to "who orchestrates the others best," and it claims frontier-level results without the frontier price tag.

In its annual filing, Oracle said it shed about 13% of its workforce over the past year, down to roughly 141,000 people, and tied the cuts directly to deploying AI across the company. In the same stretch, its spending on data centers and equipment jumped 162% to $55.7 billion. Why it matters: this is the part of the AI story that never makes the keynote. The boom is expensive, and even a cash-rich incumbent is trimming payroll to fund the concrete, chips, and power. When a company frames layoffs as both an AI efficiency win and an AI funding move in the same breath, that's the two-sided face of this whole cycle.

A machine called LineShine, in Shenzhen, topped the global Top500 ranking at 2.2 exaflops, meaning more than two quintillion math operations a second, roughly 20% faster than the US champion it dethroned. The headline isn't the speed, it's the parts list: homegrown processors, a homegrown interconnect (the wiring that ties thousands of chips into one machine), and a homegrown operating system. Why it matters: it's the clearest proof yet that export controls are pushing China to build the entire stack itself, which over time points to a world with two separate chip ecosystems instead of one.

Quantum Computing Inc. closed its purchase of NHanced Semiconductors for $73 million up front, with up to $72 million more if performance targets are hit, and said it will fast-track a second plant it calls Fab 2, the facility where chips get physically built. The deal brings advanced packaging and photonics manufacturing in-house, the step that turns a lab prototype into something you can actually ship. Why it matters: owning the factory is how a research-stage company proves it can produce at scale, which is the difference between a science project and a real supplier.

The phone-chip giant is negotiating to acquire Tenstorrent, an AI-chip startup run by legendary designer Jim Keller that builds on RISC-V, an open, royalty-free chip blueprint anyone can use instead of paying to license the usual designs. Why it matters: Qualcomm's smartphone growth is slowing, and this would buy it a real seat at the AI-data-center table that Nvidia and AMD dominate. The catch: $10 billion is a steep price for a bet that could still fall apart.

The company posts results after the market closes, with Wall Street expecting record profit margins near 81% and revenue up sharply on AI demand. Days earlier, Micron signed a deal to supply Anthropic, the maker of Claude AI, with memory and storage, and it even took a stake in Anthropic's latest funding round. The key term is HBM, high-bandwidth memory, the ultra-fast memory bolted right next to an AI chip so the chip never waits on data. Why it matters: memory used to be a boom-and-bust commodity, and now it's a sold-out, must-have input, with Micron's entire capacity booked through 2026.

Elon Musk's rocket company signed a deal worth up to $6.3 billion with an open-source AI startup called Reflection AI, which will pay about $150 million a month for access to Nvidia chips at SpaceX's "Colossus 2" data center. Why it matters: SpaceX, a rockets-and-satellites company, has quietly turned its data center into a cloud business, joining a roster of compute landlords that already includes deals with Anthropic and Google. The twist worth noting: Nvidia put $800 million into Reflection and also supplies the chips it will run on, so Nvidia sits on both sides of the same trade.