Compute

Kaohsiung marked the park on 21 September. Reuters attaches two TSMC buildings and a Q4 2029 start that the city note does not repeat.

Kioxia and Sandisk plan more than $31 billion of Japanese flash-memory investment through 2032 across Yokkaichi and Kitakami. The spending is explicitly contingent on government support, with no subsidy amount specified.

The inference ASIC posts higher peak work per watt and lower end-to-end latency against GB200/GB300 baselines in OpenAI’s own tests. Deployment inside its infrastructure is planned by year-end; foundry, yield and volume remain unpublished.

DeepSeek’s V4 tariff crossed its effective clock with seven peak hours and 17 off-peak hours. A separate vLLM merge tuned V4 decode for AMD’s gfx950, giving operators two different levers on the same day.

The neocloud reports 1.5 gigawatts active against 3.7 gigawatts contracted, while debt reaches $35.1 billion. Demand is booked; the costly part is switching it on.

Texas is making large-load applicants back megawatt requests with cash, while a published proposal would leave 80% of residual security unrecoverable in specified exits. Data centres dominate a queue whose stated size dwarfs the grid’s record peak.

Google has broken ground on a $15 billion, gigawatt-scale AI hub. Clearances cover Tarluvada and Rambilli while a third campus faces a wildlife-boundary challenge, and the cooling-water budget remains unpublished.

Europe has invited bids for up to seven AI gigafactories, with awards due in early 2027 and operations targeted within 18 months of signature. The €30 billion headline remains a funding target whose chips, grid works, sites and private capital still have to be assembled.

Chinese military-linked researchers are turning U.S. model outputs into training material for smaller domestic systems aimed at secure networks, disconnected UAVs and tactical processors. The papers map a chain from API to hardware while leaving deployment, accuracy and compute dependence unproven.

vLLM has merged Kimi K3 into mainline, while SGLang, llama.cpp and community quantizers occupy separate runnable, proposed and loadability states. [E1][E2][E3][E6][E7]

Moonshot put 1.56 terabytes of Kimi K3 weights on Hugging Face. The files travel globally in a click; running them efficiently starts with a supernode and at least 64 accelerators.

A disturbance centered on Northern Virginia abruptly removed about 3% of the system’s demand. Data centers are the leading suspect, but the operator has not established the cause.

Twenty-nine governments signed onto a Shanghai-based cooperation body as Beijing pitched wider access to models and standards. New York and Australia, meanwhile, treated compute as infrastructure subject to domestic terms.

A one-year pause on projects drawing 50 megawatts or more puts land, electricity, taxation and local consent inside the AI buildout’s cost base. Compute has entered state politics.

A record quarter arrived with another vast commitment to U.S. manufacturing. The investment points to durable demand for advanced computing while leaving timing and utilization contingent on future markets.

Agility Robotics is going public through a $2.5 billion SPAC, the first pure-play humanoid maker on a US exchange, while Unitree has filed to list in Shanghai. The robots are still learning to walk a shift; their makers are already learning to raise from the public.

SK Hynix opened at $170 and closed near $168, up about 13% from its $149 offer, valuing the memory maker close to a trillion dollars. The desk had called a hold at or above the offer; the market answered louder, and the company says the chip shortage behind it only deepens.

A humanoid works a gate at Haneda, Figure units log hours on BMW's line, and a robot fleet just finished a 100-megawatt solar farm. Bank of America sees shipments jumping from tens of thousands this year toward millions by 2030. The demos are becoming shifts.

SK Hynix priced its US listing at $149 a share and raised about $26.5 billion, the largest ever by a foreign company on Wall Street. It began trading in New York today; the bet underneath is the high-bandwidth memory bolted to almost every AI accelerator.

UBTech calls its new UWORLD U1 the first full-size humanoid built for mass production and logged 13,361 orders on launch day; its industrial Walker S2 is already shipping. Bank of America expects the fleet to climb from thousands now to millions this decade. What limits that ramp sits upstream, in memory and power.

SK Hynix set the price for its Nasdaq debut after orders ran more than seven times the shares on offer, sizing the raise near $28 billion. It would rank among the largest share sales ever, and it rests on one bet: that AI's hunger for high-bandwidth memory holds.

Seoul and Washington are steering billions toward power-rich land, permitted generation, and domestic chips while China's low-cost open models narrow the frontier gap. Semiconductor stocks fell on 2 July as investors weighed megawatt megaprojects against cheaper inference anyone can download.

Bloomberg reported that Anthropic accused operators linked to Alibaba of running a large distillation campaign against its commercial systems. Chinese developers are still releasing competitive open-weight offerings at lower cost, and enterprise buyers are routing spend toward cheaper inference, according to the New York Times and CNBC.

Sakana AI’s orchestrator turns a pool of models into one API. Its scores are the clean part. Daily use is where the claim gets harder.

AI capacity is being built as physical infrastructure: powered, sited, financed and politically prioritized. Chips still matter, but the binding constraint is moving toward electrons, memory capacity and software that lets compute move across silicon.