DeepSeek open-sourced DeepInfer, the inference framework that powers its production AI services. It reports that one cluster serves 155 trillion tokens per day and that off-peak pricing comes to $38 per GPU hour. DeepInfer is now available under an Apache 2.0 license on GitHub with support for DeepSeek-V4.1-Flash, while some performance-critical components tied to internal infrastructure and non-Nvidia determinism remain unreleased.
SpaceX reported higher-than-expected profitability driven by its AI-focused “Super Intelligence” business as it increased compute and broadened its data-center and Grok offerings. The company posted $14.6 billion in third-quarter revenue and $4.4 billion in net income. The results push SpaceX back into quarterly profitability while it continues heavy capex at $19.6 billion to expand data centers and Starship, and Musk forecasts 2027 annual recurring revenue between $500 billion and $1 trillion.
SF Compute’s CEO Evan Conrad described how GPU compute trading depends on physically settling contracts by running the underlying data-center “hotel” rather than only paying against an index. He said 5% of SF Compute’s total capacity is currently placed on the order book to make the market liquid. As more capacity is moved onto the order book and additional supply is brought online, compute contracts become more tradable despite changing GPU and software constraints.
Google announced TorchTPU, its open-source PyTorch variant for running AI apps on its TPUs and Google Cloud without rewriting. It said PyTorch is used by 85% of AI apps. As a result, developers can port existing PyTorch apps to GPUs-optional hardware like TPUs, intensifying pressure on Nvidia’s chip dominance.
HPE introduced HPE Private Cloud AI as a production-ready enterprise platform for operating private AI with configurable efficiency controls. The offer includes a 90-day engagement with forward-deployed engineers to help deploy initial use cases and set repeatable operating practices. This shifts enterprise AI management from tracking token or GPU-hour costs toward measuring effective cost per business-relevant outcome by improving accelerator utilization and routing across models and workloads.
DeepSeek open-sourced DeepInfer, the inference engine behind its production API, after earlier saying the full approach was not yet practical for public use. The release, dated October 20, 2026, is available under the Apache 2.0 license and is positioned to run DeepSeek’s V4.1 Flash model on developers’ own infrastructure. Developers can now study and adapt DeepSeek’s inference optimizations (and also see day-one support for Nvidia SM100 and Huawei Ascend 950 NPU platforms), though some production-serving components are still missing.
Daytona, a startup focused on virtual machines for AI agents, raised a $60 million Series B and reached a $1 billion valuation. The round values the company at $1 billion. Daytona plans to hire go-to-market sales staff and engineers and aims to scale faster sandbox provisioning with 60-millisecond setup for agent “computers,” addressing the speed limits of traditional cloud VM spin-up.
Daytona raised a $60 million Series B to support cloud infrastructure for AI agents at a $1 billion valuation. The round valued the company at $1 billion and brings total funding to $91 million. Daytona plans to expand its agent cloud platform beyond its sandbox, scale infrastructure, and grow its international team.
llm-d contributors announced llm-d 1.0 at PyTorch Conference North America as a Kubernetes-native distributed inference serving network intended for production deployments. The release highlights that its prefix-aware router captured 97% of available cache reuse in an agentic coding workload test. It adds operational maturity for “day 2” needs like multi-tenant quality of service, zero-downtime upgrades, and sustained efficiency under changing traffic and failures.
Lithography exports—specifically DUV immersion lithography systems—are central leverage in shaping China’s ability to scale AI chipmaking and compute. China imported about 90 immersion lithography systems per year in 2024–25, nearly two-thirds of ASML’s immersion exports. Congress’s proposed MATCH Act would restrict DUV immersion exports and servicing to reduce that leverage loss and buy time for US AI compute advantage into the 2030s.
Nebius and Nscale are planning data-center infrastructure built around NVIDIA Vera CPUs to support agentic computing. Vera targets high concurrency, sustaining speed across thousands of tasks running simultaneously. This shifts European AI infrastructure toward more CPU-driven retrieval, tool execution, code running, and sandbox throughput alongside GPUs for model workloads.
Nvidia invested in four German startups across robotics, healthtech, audio, and space technology. Nvidia’s European pledge totals nearly $10bn. The investments add more European startup funding and support Nvidia’s push to drive adoption of AI models in industrial and physical-AI applications, including partnerships and a Germany joint venture.
Samsung is expanding its Pyeongtaek, South Korea semiconductor complex to make more memory stacks for GPUs used in generative AI workloads. Construction of a fifth memory fab was halted during a 2024 demand drop but restarted after AI-driven demand surged. The expanded build-out is meant to help Samsung regain market position in memory and keep fab construction operating through rising demand expected to continue through 2028.
Barcelona Supercomputing Center (BSC) put its AI Factory Accelerator Programme into full operation and integrated access to some BSC AI Factory services into NVIDIA’s Inception portal. Since its April 2025 launch, BSC AI Factory has delivered more than 800 services to private and public-sector clients. Startups now get sovereign compute plus support and training through both programs, and an extension of MareNostrum 5 will further expand access via AI Factories across Europe.
Hugging Face released TorchTPU, adding a native TPU backend to PyTorch so Transformers and Diffusers can run on TPUs by selecting the device. It reports 96.5% of the Transformers TPU test suite passing, with 34 issues found and 19 fixed so far. TPU support becomes a first-class PyTorch workflow (device_map="tpu" / pipe.to("tpu")) without needing torch_xla or Optimum TPU.
PyTorch 2.15 upgrades its MPS backend for macOS so it can run more ops on dedicated Metal kernels and adds FP8 support for Mac acceleration. FP8 tensors on MPS use one byte per value, with a float8_e4m3fn format that stores in half the space of BF16. As a result, local LLM decoding is faster (including MoE and softmax/linear layer paths), video VAE decoding needs far less driver memory, and macOS 15+ plus Python 3.11+ become required.
Every AI story that matters,
in your inbox by 8am.
TLDRocket reads all relevant sources, removes duplicate coverage, and summarises the
day in two minutes. Follow companies and topics for alerts, or get the
briefing in Slack. Free, no spam, unsubscribe anytime.
Reading TLDRocket needs no cookies, and the readership counts we rely on come from
our own cookieless analytics. Google Analytics is the exception: it sets cookies and
reports to Google, so it stays switched off until you allow it. You can change your
mind any time from “Cookie settings” in the footer.
Strictly necessary
Session security and form protection (tldrocket-session,
XSRF-TOKEN, 2 hours). The site cannot work without them,
so they need no consent.
Always on
Google Analytics 4 (_ga,
_ga_<id>, up to 2 years). Measures which
stories and sections readers use. Google acts as a third-party processor and may
store the data outside the EU. No advertising, no profiling, no data sold.
