Capital & Industrial Strategy

Bloomberg reports Anthropic is considering going public as soon as October 2026, racing OpenAI to public markets. SoftBank's new $40 billion loan from JPMorgan and Goldman Sachs — unsecured, 12-month term — appears structured specifically to bridge OpenAI to an IPO, according to TechCrunch. The loan allows SoftBank to maintain liquidity for OpenAI investments without diluting equity positions before a public listing. Meanwhile, Anthropic won a court order blocking its designation as a supply chain risk by the U.S. government, though Politico notes lawyers say the victory may prove temporary pending D.C. Circuit appeals under Trump-appointed judges.

The timing convergence suggests the venture window for foundation models has effectively closed. Both companies need public markets to provide liquidity to employees and early investors, and to fund continued compute expansion without accepting increasingly onerous terms from strategic investors. Anthropic's regulatory challenge adds urgency — an IPO would complicate government attempts to restrict the company, as public shareholders create political obstacles to post-listing intervention.

Bloomberg reports Physical Intelligence, founded by AI academics and former Google DeepMind researchers just two years ago, is negotiating a $1 billion round at an $11 billion post-money valuation. This represents one of the largest funding rounds for an AI robotics company and signals investor belief that the next wave of AI commercialisation lies in physical applications rather than purely digital foundation models.

The valuation appears driven by Physical Intelligence's focus on building general-purpose robotic control systems — essentially applying foundation model concepts to physical manipulation tasks. The company's pedigree matters: DeepMind alumni carry credibility that translates directly into fundraising leverage. The round size and valuation suggest investors see robotics as offering clearer paths to revenue than open-ended AI research, with manufacturing, logistics, and warehouse automation providing immediate enterprise customers willing to pay for productivity gains.

Bloomberg reports Meta is paying for construction of seven new natural gas-fired plants to power its Louisiana Hyperion data center, the company's most power-intensive facility. The move marks a strategic shift: rather than waiting for utilities to expand grid capacity, Meta is directly financing generation infrastructure. Separately, Bloomberg reports Microsoft acquired 900MW of data center capacity originally developed for Oracle and OpenAI after both companies walked away from the Texas project. Financial Times adds that Google is nearing a deal to help finance a multibillion-dollar data center leased to Anthropic, with the Texas site designed to avoid grid connection delays through direct gas supplies.

These moves reveal how power availability has become the binding constraint on AI infrastructure deployment. Hyperscalers are no longer willing to let utility planning cycles dictate their expansion timelines. By financing generation directly, they gain scheduling certainty and can accelerate deployment, but they also assume long-term commodity price risk and regulatory exposure. The OpenAI-to-Microsoft capacity transfer suggests even well-funded AI companies are struggling to execute infrastructure builds at the pace their compute roadmaps require.

Financial Times reports memory chip stocks shed $100 billion in market capitalisation after new research indicated AI data centers will require significantly less memory than investors had anticipated. The sell-off unwinds a trade that had propelled memory chip manufacturers to record valuations based on assumptions that AI inference workloads would drive explosive DRAM demand. The research suggests modern AI architectures rely more heavily on high-bandwidth memory directly attached to accelerators rather than traditional DRAM pools, and that inference optimisation techniques reduce memory footprints more than earlier projections assumed.

The correction reveals how speculative positioning in AI infrastructure plays had disconnected from technical fundamentals. Investors treating memory chips as a pure AI proxy failed to account for architectural evolution and software optimisation. The speed of the unwind — $100 billion in a single trading session — indicates institutional positioning was crowded and leveraged.

Reuters reports Huawei's new AI chip is attracting orders from ByteDance and Alibaba, indicating China's largest technology companies view domestic alternatives as viable despite performance gaps with Nvidia hardware. This follows Reuters reporting that Chinese universities with military links purchased Super Micro servers containing restricted AI chips, demonstrating continued leakage in U.S. export controls.

The willingness of ByteDance and Alibaba to standardise on Huawei chips represents a strategic shift. These companies previously absorbed cost and performance penalties to maintain access to Nvidia hardware through grey market channels. Committing to Huawei suggests either that performance gaps have narrowed sufficiently for production workloads, or that they assess U.S. restrictions will tighten further and are securing domestic supply chains while still possible. Either interpretation indicates China's AI ecosystem is decoupling faster than U.S. policy architects expected.