Yandex confirms 2.84% of cloud customer data is gone for good after drone strikes
After October 8–11 drone strikes on three data centers, Yandex restored 97.16% of affected customer data; 2.84% is unrecoverable. New VMs default to backups on.
Published entries filed under “Compute” in News, newest first by publication date on this site.
46 entries
After October 8–11 drone strikes on three data centers, Yandex restored 97.16% of affected customer data; 2.84% is unrecoverable. New VMs default to backups on.
Microsoft 365 broke from Exchange Online outward on November 12, hitting Outlook, Teams and Copilot; recovery came November 14, blamed on a capacity upgrade.
AMD announced its Zen 6 EPYC Embedded 9006 series on November 10, with up to 256 cores, PCIe Gen 6 and CXL 3.1; the first model is scheduled for November 15.
Google’s blog dated TPU 8tx and 8i Max to Nov. 10; two media reports are dated Nov. 9. One targets confidential inference; the other pairs two chips.
The FT reports Anthropic agreed to pay $3.4 billion to take over OnePWR's 1.7GW Caterpillar generator order — the largest power-equipment purchase by an AI lab.
Tesla's November 6 slides lay out the SI roadmap: SI6 claimed at 80x HW4 compute for 2029-2030; Dojo D3 returns with chiplets plus stacked HBM for space.
The Raspberry Pi 5 16GB now costs $385, up $80, more than triple its $120 launch price; Raspberry Pi blames a lack of alternative 16GB DRAM supply.
Diamond Foundry says it grew the first 200mm single-crystal diamond wafer, aimed at AI chip cooling; no third-party validation yet.
SpaceX subsidiary MZX reportedly offered to buy 418 homes near its Colossus AI data center in Southaven, Mississippi, if owners withdraw from a noise lawsuit.
Linux on an RTX Spark laptop runs GPT-OSS-120B at 13.9 tok/s versus Windows at 65.4 — MindStudio traces the gap to an encrypted firmware power table.
An eBay seller is asking $59.7M for a decommissioned AI supercomputer: 300 racks, 9,600 A100 40GB GPUs, 4,200 EPYC CPUs — about double the parts value.
nv-uefi-unlocker boots from a stick and unlocks capped RTX Tensor Cores before the driver loads — a 4060 Laptop gains 65% FP16, no VBIOS flash, no persistence.
Samsung says 2027 HBM supply volume is booked at target prices, with 10 LTAs signed and DRAM tight through 2028.
Announced Oct 27: Google TPUs land on OCI in calendar 2027, the first hyperscaler outside Google to offer them; ORCL rose about 5% that day.
Micron CTO Scott DeBoer called 4F² density gains "a bit of a myth" on Oct 21, pulling the design from the roadmap against Samsung, SK hynix and CXMT.
Bank of Korea's Oct 26 reply to parliament models a US AI-capex slowdown that cuts chip exports $50–100B below forecast and 2027 GDP by 0.4 points.
The open-source GPU Unlock module rewrites Ampere speed registers before NVIDIA's driver binds, lifting RTX 30 BF16 GEMM up to +80%. Linux only.
VentureBeat's VB Pulse surveys: 79% of enterprises report GPUs at half capacity or less, yet 70% plan a new or replacement AI compute platform within a year.
On the a16z podcast aired October 23, Amazon CEO Andy Jassy said AWS 2026 capacity is sold out and 2027 largely so — and the shortage will not end in 2027.
On October 21 Nvidia announced DSX POD at GTC Berlin: seven Vera Rubin racks pre-built in a container, a 14-day site claim, and early-2027 shipping.
On October 21 OpenAI put GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna into Google Cloud's model garden in private preview, billable against GCP commitments.
Google open-sourced TorchTPU under Apache 2.0 on October 20. Google says PyTorch code needs minimal changes to run on TPUs. For teams cutting Nvidia reliance.
Humming is an Apache-2.0 quantized kernel library from vLLM and Ant Group. The project says DeepSeek-V4.1-Flash output throughput rose 14–43% on eight H20 GPUs.
llm-d 1.0 shipped October 20 with multi-tenant QoS and zero-downtime upgrades for self-hosted inference. Disaggregated setups need LeaderWorkerSet v0.11.1+.
DeepSeek open-sourced DeepInfer under Apache-2.0, the framework behind its V4.1 Flash service, for NVIDIA SM100 and Ascend 950. KV cache code is not included.
On 20 October, a power event in one Google Cloud London zone left 38 products affected; Revolut and England and Wales court video links reported errors.
SpacemiT launched the 46-core RISC-V server CPU V100 on October 19: eight DDR5 channels, 120W and a full server-spec compliance claim. No price or ship date.
Meta says Prometheus, its supercluster in New Albany, Ohio, finished phase 1 and trained Muse Wonder; the target is about 1 million GPUs.
DGX Spark local models: n8n's October 19 tutorial runs seven models on one box, 120 tok/s for one user. Its LiteLLM 1.84.0 floor lets affected versions through.
Leaker Jaykihn claims Intel's next edge Nova Lake Xeon E ships single core types only — up to 16 P-cores or 32 E-cores, no LPE. Intel has not confirmed it.
Nvidia's Jeff Fisher said the RTX 5090 is not discontinued and will keep shipping, while conceding tight supply; third-party US listings run near $8,500.
UBI Research sees memory near 40% of flagship phone BOM by late 2026, six times the display; Counterpoint measured 42% in Q2. Makers are cutting screen specs.
OpenAI compute chief Sachin Katti leaves Nov 15 for a new venture, staying as an advisor; ex-xAI infra head Uday Ruddarraju takes over compute.
Microsoft announced its Maia 300 inference chip on October 12: 40% more tokens per megawatt than leading GPU systems, and multi-gigawatt deployments from 2027.
Google and Constellation signed a 20-year PPA on October 6: $4.3B to uprate 11 reactors for 890 MW, plus a separate 15-year, 2,700 MW supply deal.
A Planet-built prototype rides Transporter-18 to orbit; the Joule paper says scale needs 370,000 tons of Starship payload, about 1,800 launches.
Micron says supply-demand is "only getting tighter": 75% of 2027 output committed, talks cover 2028, and HBM takes 30% of DRAM capacity next year.
OpenAI and Synopsys announced a multi-year deal on September 30 to co-develop GPT-Synopsys, a chip-design model trained with licensed EDA tools.
Google open-sourced AX under Apache 2.0: agents as stateful long-running actors, sub-second checkpointing, dense per-host multiplexing, on top of Kubernetes.
DeepSeek disclosed DSec, the sandbox infra behind its agent training: four execution backends, on-demand images, about 3 million sandboxes a day.
DeepSeek published six Ascend component repos on September 30; it says every V4-series NVIDIA operator now has an Ascend counterpart.
OpenAI will run managed agents on AWS and bring ChatGPT to Slack and Teams without individual licenses, with a private-inference preview coming this fall.
Per a Sept 28 filing, CXMT will put RMB 24.1B into a technology R&D project and RMB 10.8B into its Hefei memory test base Phase II, using over-raised funds.
Per The Information, China’s MIIT asked ByteDance and Alibaba to report RTX Pro 5500 purchase plans and signaled approval; it may skirt US export controls.
Akamai announced on September 24, 2026, that Anthropic will spend $11.6 billion over seven years on its cloud, the largest contract in Akamai’s history. Akamai also issued Anthropic a warrant, and the deal could grow to about $20 billion.
Inferact, founded by the original vLLM team, wrote a megakernel inference kernel for Google TPUs. Paired with DeepSeek’s DSpark speculative decoding, 16 TPU v7 chips served Kimi K3 at 709 tokens per second versus 452 on GB200 under the same setup, QbitAI reports. The code is open source.