AI & Compute Infrastructure — August 13, 2026

🇺🇸Databricks closed $5 billion in financing at a $190 billion post-money valuation for enterprise AI capabilities. #

Databricks announced the completion of a $5 billion financing round at a $190 billion post-money valuation. The source characterized Databricks as a privately held data-analytics software company and said its annualized revenue had exceeded $7 billion, with second-quarter revenue growing more than 80% year over year.

The company said the proceeds would strengthen its enterprise artificial-intelligence capabilities. The named areas include Unity AI Gateway, a governance and control tool, and Genie, its intelligent-agent product.

The transaction followed another financing completed six months earlier. That earlier capital package consisted of $5 billion in equity financing and $2 billion of additional debt capacity at a reported valuation of $134 billion; the source does not provide the new round’s security type, investor allocation, liquidation terms, collateral, or dilution.

• Databricks

🇺🇸Anthropic’s CFO began early IPO investor meetings; discussions have not yet covered valuation. #

Anthropic Chief Financial Officer Krishna Rao was reported to be leading preliminary meetings with potential initial public offering investors. The discussions remained at an early, high-level stage and had not addressed a proposed IPO valuation.

According to the more detailed report, the meetings covered Anthropic’s Claude model family, Claude Code programming assistant, enterprise-market positioning, management team, and product development. Specific financial information was also reportedly not yet part of the discussions.

The report said Anthropic had confidentially submitted listing documents to the U.S. Securities and Exchange Commission in June but had not announced a formal IPO timetable. It also said some investors expected a possible $2 trillion IPO valuation, while explicitly noting that this was not an official company target.

• Anthropic’s CFO

🇺🇸AMD priced a four-part $4.75 billion senior-notes offering, spanning maturities from three to ten years. #

Advanced Micro Devices priced an aggregate $4.75 billion senior-notes offering in four parts, according to a message identifying an SEC filing as its source. This is a priced debt offering rather than the earlier reported maximum of as much as $5 billion.

Prior reporting said AMD planned four note tranches with maturities ranging from three to ten years. The longest-dated tranche initially carried price guidance of approximately 1.15 percentage points above the corresponding U.S. Treasury yield, but the supplied sources do not provide its final spread.

The available source material does not state the final coupon for each tranche, use of proceeds, collateral, financial covenants, or whether the notes are guaranteed. It also does not identify any equity-conversion feature, so the financing should be treated only as the disclosed senior-note issuance.

• AMD

🇺🇸The United States imposed 100% tariffs on sensitive drones and 25% tariffs on smaller models, effective after 21 days. #

President Donald Trump signed a proclamation imposing ad valorem tariffs of 10% to 100% on imported drones and components, citing national-security threats. The 100% rate covers drones of specified sizes or sensitive capabilities, including drones with a maximum takeoff weight above 25 kilograms and drones equipped with thermal-imaging capabilities.

A 25% tariff applies to certain smaller drones that lack the specified sensitive capabilities, as well as some other drone components. Drones and components from the European Union, Japan, Liechtenstein, South Korea, and Switzerland are subject to a 15% rate under the announced framework.

Drones from the United Kingdom are subject to a 10% tariff if their hardware, software, and technology are substantially sourced from the listed countries and regions or the United States. The principal tariffs take effect 21 days after signing, while tariffs on drone components not classified as highly sensitive begin after 180 days.

• United States

🇺🇸New York opened a $60 million request for proposals for up to four regional quantum-commercialization hubs. #

New York Governor Kathy Hochul announced a competitive request for proposals to establish as many as four Regional Quantum Technology Commercialization Hubs across the state. The initiative received $60 million in funding through New York’s fiscal-year 2027 budget.

The hub program was first proposed in Hochul’s 2026 State of the State agenda. Its stated purpose is to help translate quantum-technology research into commercial activity through regional centers.

The announcement constitutes an open competitive funding process, not completed awards to four recipients. The supplied source does not identify selected operators, individual hub allocations, matching-fund requirements, or project-completion dates.

• New York

Lenovo reported RMB 360 billion of AI-server backlog, up from RMB 140 billion the previous quarter. #

Lenovo Group (联想集团) said its AI-server backlog reached RMB 360 billion, up from RMB 140 billion in the preceding quarter. Chairman and CEO Yang Yuanqing said AI demand was only beginning and that tight memory supply would continue through both this year and next year.

Lenovo’s Infrastructure Solutions Group generated first-quarter revenue of RMB 57.9 billion for the period ended June 2026, an increase of 98% year over year. Operating profit was approximately RMB 5.3 billion, also up 98% year over year.

The infrastructure division’s operating margin was 9.1%. That represented an improvement of 11.1 percentage points from a year earlier and 5.5 percentage points from the preceding quarter.

The RMB 360 billion figure is described as reserved AI-server orders rather than recognized revenue or completed server shipments. The source does not disclose the backlog’s accelerator mix, customer concentration, cancellation provisions, delivery schedule, or associated power and data-center capacity.

• Lenovo

SMIC reported second-quarter revenue above $3 billion, with shipments rising 14.4% sequentially and utilization reaching 93.7%. #

Semiconductor Manufacturing International Corporation, or SMIC (中芯国际), reported second-quarter sales of approximately $3 billion, up 20% sequentially and 36.1% year over year. Its gross margin was 25.3%, compared with 20.1% in the first quarter and 20.4% in the second quarter of 2025.

Co-CEO Zhao Haijun said wafer shipments increased 14.4% from the previous quarter, while average wafer selling price rose 5.7%. The company attributed higher shipments principally to sharply increased demand for AI-supporting chips and customers pulling forward shipments.

Capacity utilization reached 93.7%, an increase of 0.6 percentage points sequentially. SMIC added the equivalent of 8,000 12-inch wafers per month of capacity during the quarter.

China, the United States, and Eurasia accounted for 90%, 8%, and 2% of quarterly sales respectively. Twelve-inch and eight-inch wafers represented 78% and 22% of wafer revenue; revenue from the two categories increased 24% and 11% sequentially, respectively.

• SMIC

Lam Research plans to invest over $3 billion across five years to expand global laboratory capacity by more than 50%. #

Lam Research disclosed plans to invest more than $3 billion over the next five years in its global laboratory network. The expenditure is intended to expand the company’s laboratory capacity by more than 50%.

The stated scope is laboratory capacity rather than semiconductor-wafer fabrication capacity, data-center capacity, or GPU production. The source describes a global network expansion but does not provide a country-by-country allocation.

The announcement is a forward investment plan rather than completed capital expenditure. The supplied source does not specify annual spending levels, individual facility schedules, funding arrangements, or the exact laboratory-capacity measurement behind the targeted increase.

• Lam Research

Sandisk taped out its first high-bandwidth flash chip and expects samples next year; Meta joined the open-standard group. #

Sandisk said it had completed tape-out of its first high-bandwidth flash, or HBF, chip. Tape-out refers here to completion of the chip design for manufacturing, not volume production or customer qualification.

A Sandisk executive said the company expects to provide the first HBF samples for inference equipment to customers next year. The source does not state a firm sampling quarter, production volume, bandwidth, capacity, process node, or commercial-shipment date.

Meta Platforms will join the industry group developing an open HBF standard. The announcement identifies Meta’s participation in standards development but does not disclose a purchase agreement, supply commitment, or deployment volume.

• Sandisk
• Meta

Apple’s Houston facility shipped its first advanced AI servers and will begin Mac mini production later this year. #

Apple opened an advanced manufacturing center in Houston and said the facility had shipped its first advanced AI servers. The company has invested hundreds of millions of dollars in the site, according to the supplied report.

Parts of the Houston facility were already assembling AI servers for Apple’s own use. The sources do not identify the servers’ processor or accelerator configuration, shipment quantity, rack architecture, power requirements, or destination.

Apple also said Mac mini desktop production would begin at the Houston facility later in 2026. CEO Tim Cook and U.S. Commerce Secretary Howard Lutnick attended the facility’s unveiling.

• Apple’s Houston facility
• Mac mini

Mitsubishi Power’s large-frame gas-turbine backlog reached 35 gigawatts; the company plans to double manufacturing capacity. #

Mitsubishi’s backlog for large-frame gas turbines reached 35 gigawatts. The figure measures the generating capacity represented by turbine orders, not currently operating generation or completed power plants.

The company plans to double its manufacturing capacity. Chief Financial Officer Hiroshi Nishio said Mitsubishi continued to see a strong supply-demand environment for gas turbines in the United States.

The supplied source does not state the baseline manufacturing capacity, the date by which the doubling will be completed, or how much of the 35-gigawatt backlog relates specifically to data centers. It also does not provide customer concentration, turbine delivery schedules, or project-financing details.

Jereh received an initial prepayment under a $1.465 billion gas-turbine generator order from an international cloud provider. #

Jereh Group (杰瑞股份, 002353.SZ) said its controlled subsidiary J&F Power Systems LLC signed a gas-turbine generator-set supply contract with an unnamed internationally known cloud-service provider in July. The customer subsequently issued a $1.465 billion purchase order.

Jereh said it had received the first prepayment under the order. Receipt of an initial prepayment marks progress beyond an unsigned plan or memorandum, but the source does not state the prepayment amount, remaining payment milestones, or whether all conditions for the full order have been satisfied.

Separately, Jereh reported RMB 10.046 billion of new orders in the first half, up 1.67% year over year. Its outstanding order book at the end of the reporting period was RMB 13.995 billion, an increase of 13% from a year earlier.

The source does not disclose the number or model of gas-turbine generator sets, total electrical capacity, deployment location, collateral, delivery timetable, or the cloud customer’s identity. It also does not state how much of the $1.465 billion has been recognized as revenue.

• Jereh

Open Access Technology International seeks federal funding for software it says could unlock up to 20% more bulk-grid capacity. #

Open Access Technology International, or OATI, is seeking federal funding for broad deployment of a software-based grid initiative. OATI says the initiative could unlock as much as 20% more capacity on the bulk power system.

The proposed components include dynamic line-rating software, near-real-time coordination across regions, and artificial-intelligence-enhanced dispatch of power resources. Dynamic line ratings are presented as part of the software package rather than construction of new transmission circuits.

OATI says the initiative would not require new poles or wires. The 20% figure is the company’s stated potential, while the source does not report a completed federal award, deployed system-wide capacity increase, or independently measured result.

• Open Access Technology International

Alibaba Cloud’s Wulanchabu data center uses roughly 90% green electricity; modular construction reduced delivery to 100 days. #

Alibaba Cloud (阿里云) said its Wulanchabu data center is one of its five hyperscale data centers. Approximately 90% of the electricity used at the facility is described as green electricity, providing a lower-carbon operating environment for high-density computing.

The company uses a fully modular design and construction architecture at the location. It said this approach shortened the delivery period for a large artificial-intelligence data center to 100 days.

Wulanchabu was also the initial launch site for Alibaba Cloud’s Lingjun Zhenwu M890 supernode instance (灵骏真武 M890 超节点实例), which became available on Aug. 12. The source does not state the facility’s total megawatt capacity, accelerator count, rack density, power-usage effectiveness, or the accounting basis for the 90% green-electricity share.

A federal geothermal parcel in New Mexico sold for a record $701 per acre ahead of Utah’s August lease sale. #

A federal geothermal parcel in New Mexico sold in June for $701 per acre. Utility Dive described the result as a record-setting bid for a federal geothermal lease.

Analysts cited stronger competition among bidders as the Bureau of Land Management prepared another geothermal lease sale in Utah. That Utah sale was scheduled for Aug. 18.

The $701-per-acre figure is a lease-sale bid, not the construction cost, electricity price, or value of an operating geothermal power plant. The supplied source does not state the parcel’s acreage, bidder identity, expected generating capacity, exploration schedule, or development commitments.

• New Mexico
• Utah’s August lease sale

OpenAI partnered with Cerebras for GPT-5.6 Sol UltraFast inference, reaching up to 750 tokens per second for initial customers. #

OpenAI partnered with Cerebras to provide low-latency inference for GPT-5.6 Sol through an UltraFast mode. The service was released as a limited preview to an initial group of customers rather than as a general-availability product.

OpenAI said the UltraFast tier can run GPT-5.6 Sol at speeds of up to 750 tokens per second. Another announcement described the preview as delivering speeds of up to 14 times the comparison baseline, although the source does not define that baseline.

The stated performance is an “up to” inference-throughput figure, not a guaranteed speed for every prompt, model context, or customer workload. The sources do not disclose pricing, Cerebras system counts, data-center locations, service-level commitments, or the size of the initial customer group.

• OpenAI
• Cerebras
• GPT-5.6 Sol UltraFast

Oracle Exadata Database Service became generally available on Exascale Infrastructure through Oracle AI Database@AWS. #

Oracle announced the general availability of Oracle Exadata Database Service on Exascale Infrastructure through Oracle AI Database@AWS. General availability indicates that the service has moved beyond a preview stage.

Oracle said the offering brings Exadata-class database performance and pay-per-use economics to Oracle AI Database workloads of any scale. The announcement concerns managed database infrastructure available through Amazon Web Services, rather than AWS-native database hardware.

Oracle and Amazon Web Services also signed an expanded, long-term strategic collaboration agreement. The stated focus is accelerating their Oracle AI Database collaboration, although the supplied source does not disclose contract value, minimum cloud-spending commitments, regions, or customer migration volumes.

• Oracle Exadata Database Service
• Exascale Infrastructure
• Oracle AI Database@AWS

Cascadia, an open-source distributed inference runtime, pools Intel-powered machines to run models exceeding any single machine’s capacity. #

Community Labs launched Cascadia as an open-source distributed AI-inference runtime for Intel hardware. The software was developed in collaboration with Intel.

Cascadia pools compute and memory resources from multiple Intel-powered machines. This allows the combined system to run models larger than the capacity of any one participating machine.

The announcement concerns distributed inference rather than model training or a new Intel processor. The supplied source does not identify supported Intel CPU or accelerator generations, network requirements, maximum cluster size, benchmark results, licensing terms, or production deployments.

• Cascadia
• Intel-powered machines

AMD signed a memorandum of understanding to provide £5 million in compute for Oxford-led open frontier-AI research. #

AMD signed a memorandum of understanding with the University of Oxford to support open academic frontier-artificial-intelligence research. The stated support consists of £5 million in compute.

The research will be conducted by the British Open-ended Learning and Discovery Lab, or BOLD. BOLD is a United Kingdom-based AI research laboratory led by the University of Oxford.

The transaction is at the memorandum-of-understanding stage, not described as a completed grant, equipment delivery, or binding procurement contract. The supplied source does not identify the hardware configuration, cloud provider, compute-hour quantity, deployment location, access schedule, or conditions attached to the £5 million amount.

• AMD

Google Gemini 3.7 Flash launched at introductory pricing of $0.75 input and $3.75 output per million tokens. #

Google launched Gemini 3.7 Flash with introductory pricing of $0.75 per million input tokens and $3.75 per million output tokens. The promotional prices are scheduled to remain in effect through the end of 2026.

Google said Gemini 3.7 Flash outperforms Gemini 3.6 Flash on complex-document processing. The company also emphasized improved code generation and debugging, including first-pass code intended to be closer to production-deployment requirements.

Gemini Spark began using Gemini 3.7 Flash on the launch date. Google also said the model includes additional safeguards addressing malicious attacks and misuse involving chemical, biological, radiological, and nuclear domains.

The pricing applies separately to input and output token processing and should not be combined into a single per-million-token rate. The sources do not provide post-promotional pricing, context-window size, throughput, regional availability, or service-level guarantees.

• Google Gemini 3.7 Flash

Researchers developed a GPU-based scientific-data compression method processing simulation output at 60 gigabytes per second. #

Researchers developed a graphics processing unit-based method for compressing scientific-simulation data at 60 gigabytes per second. The metric refers to the rate at which the method processes simulation output, not storage capacity or network bandwidth.

Large scientific simulations can generate terabytes of output that must subsequently be transferred, stored, and analyzed. The source notes that GPUs have increased the speed and scale of scientific computing while storage and data-transfer bandwidth have not advanced at the same rate.

The supplied description does not identify the tested GPU model, compression ratio, numerical-error tolerance, simulation type, decompression speed, or comparison baseline. It therefore supports the reported 60-gigabyte-per-second processing rate but not a broader claim about performance across all scientific workloads.

Fudan University’s Whitzard agent achieved a 91.2% real-vulnerability attack success rate across 1,507 vulnerabilities in 188 projects. #

Fudan University’s Whitzard agent (白泽智能体) recorded a 91.2% attack success rate against real vulnerabilities in CyberGym’s latest international AI-security benchmark. The result placed the system second globally and first among university teams, according to the report.

The benchmark covered 1,507 real vulnerabilities across 188 large open-source projects. Agents were required to work within complete codebases and independently perform vulnerability localization, attack construction, and validation.

The report said Whitzard ranked ahead of Anthropic’s cybersecurity-oriented Claude Mythos model and teams associated with Microsoft and Google-owned Wiz. These are benchmark standings reported by the source, not evidence of deployments against live third-party production systems.

The source characterizes the result as having been achieved at low cost but does not provide an exact training cost, inference cost, hardware configuration, attempt limit, or comparison of compute budgets. The 91.2% figure should therefore be read within the stated CyberGym benchmark scope.

• Fudan University’s Whitzard agent