Fireworks raised a $1.51B Series D at a $17.5B valuation — up from $4B nine months ago — led by Atreides Management, Index Ventures, and TCV. The inference platform says it has crossed $1B in annualized revenue, up 5x year-over-year, and now serves more than 40 trillion tokens a day, up from 15 trillion, evidence the serving layer's usage is compounding as fast as its valuation. Watch whether that growth holds as frontier labs increasingly multi-home distribution across neutral platforms rather than serve models exclusively themselves. Fireworks CEO Lin Qiao on Sequoia's Training Data podcast lays out the bet behind the round: that inference cost, not model access, is the real constraint on AI adoption, and that the future belongs to millions of specialized models rather than one dominant one.
Private Companies
Anthropic
other
Anthropic is lining up investor meetings for a public listing as soon as October, working with Goldman Sachs, Morgan Stanley, and JPMorgan, per CNBC and Bloomberg — a concrete step beyond the confidential S-1 it filed June 1 at a $965B valuation following a $65B Series H. It's the first frontier AI lab to reach this stage of the public-markets process, setting the template for how a still-unprofitable, compute-hungry business explains its economics to public investors.
Hailo
other
An unnamed Nasdaq-listed chipmaker is in advanced talks to acquire Hailo, the Israeli edge-AI chip maker once valued above $1B, in a deal reportedly worth just tens of millions of dollars, per Globes. The data center chip market has been flooded with interest, but edge is an evolving ecosystem. On-device AI and the chips needed to support it have not seen the same scale from startups.
Nscale
other
Nscale's £2B ($15B-valued) data center in Loughton, Essex, will miss its 2027 opening: the 90MW grid connection underpinning its planning approval isn't arriving in time, per Capacity Media, and the neocloud is now in talks with Bloom Energy about on-site gas fuel cells as a stopgap. It's grid interconnection showing up as project-level financial risk rather than a macro talking point, for one of Europe's most capitalized AI infrastructure buildouts.
Crusoe
partnership
Crusoe and land partner Lancium announced a 1.0-gigawatt AI data center campus on 270 acres in Childress, Texas, construction targeted for Q3 2026, using the same structure as their Abilene site: Lancium owns the land and handles power interconnection and behind-the-meter solar and storage, Crusoe designs, builds, and operates. It's another gigawatt-scale commitment stacking onto Crusoe's build-out even as the company is reportedly in talks for a $3B round at a $30B valuation.
Together
customer
Together AI and Baseten both won day-zero platform access to Inkling, the new 975B-parameter model from Mira Murati's Thinking Machines Lab, serving it via serverless, OpenAI-compatible APIs within hours of release. Landing simultaneous launch-partner status with a rival inference platform for the same model shows frontier labs now multi-homing distribution from day one rather than picking one exclusive server — a dynamic that keeps inference-layer competition on price and latency, not access.
Public Markets
TSMC beat Q2 estimates, raised FY26 revenue-growth guidance to "slightly more than 40%," lifted capex guidance to $60-64B, and committed a further $100B to Arizona — the clearest reset yet of AI capex durability, even as memory peers sold off Wednesday on China-competition fears.
Emerging
Apple has been talking to bankers and AI-chip startups about acquisitions to cut its dependence on Nvidia, after its in-house server chip ("Baltra") slipped past its 2026 target and CFO commentary dropped Apple's longstanding "net cash neutral" framing. Apple has historically avoided large deals; a shift toward buying chip-design teams would reshape the M&A market for AI silicon startups and confirm that even the most vertically integrated hyperscaler can't out-engineer Nvidia's position alone. Watch for a confirmed target.
ASML and Intel jointly announced that Intel Foundry is now shipping Core Ultra Series 3 processors on its 18A node built with High-NA EUV, the industry's first high-volume logic product on the next-generation lithography platform, with yields reportedly matching standard EUV. It de-risks the High-NA roadmap that TSMC and Samsung are also counting on for future nodes and gives Intel Foundry's turnaround story its first concrete manufacturing proof point beyond announcements.
Watch This Week
Wednesday, July 22
SK Hynix reports Q2 2026 earnings, the first full look at whether the softer-than-consensus operating-profit preview and this week's broader memory-sector selloff (Micron and SanDisk down double digits on competition from CXMT's $8.55B domestic DRAM IPO) reflect a real demand air-pocket or a valuation reset after outsized 2026 gains.