AI Cloud Upstart Volta Raises at a $2.4 Billion Valuation, with NVIDIA and Dell Joining the Round

Emerging AI cloud company **Volta Infra Holdings** has secured $300 million in venture funding along with a debt facility of up to $5 billion, at a valuation of $2.4 billion. The round was co-led by Andreessen Horowitz and Altimeter Capital, with NVIDIA and Michael Dell also participating.

The company has also reportedly landed a cloud computing contract worth $10 billion over six years with a major AI developer, though it has not disclosed the counterparty's name.

The backdrop: right now, only the largest players with solid financial foundations can afford to deploy cutting-edge AI chips. Volta — founded by veterans of Brookfield's infrastructure arm — is betting that its expertise in finance and power procurement can tear down this "capital wall." The company has already secured 1 gigawatt of power for the near term, with deployments planned in Norway (a partnership with bitcoin miner Bitdeer, 133 megawatts), Texas, and Wyoming, and it aims to have several gigawatts operational by 2030.

AI infrastructure investment has already seen a string of financing chains: Broadcom teaming up with Apollo and Blackstone for Anthropic, CoreWeave, and NVIDIA itself committing $350 billion in chip financing to OpenAI. Critics also warn that these "circular" structures, where the financiers and the customers are intertwined, carry the risk of magnified losses if AI demand falls short of expectations.

Kimi K3 at 2.8 Trillion Parameters: Safety-Stripped Versions Are Circulating

The weights of **Kimi K3**, released in late July 2026, weigh in at 2.8 trillion parameters — among the largest of any openly released model. The problem is that derivative versions with the alignment (safety mechanisms) stripped out started circulating around the time the weights were released.

An article on Qiita asks "Who is selling LLMs with their safety mechanisms removed right now?" and maps out the distribution channels and risks of these unlocked models. While open-weight releases bring greater transparency to research, the situation is forcing a fresh look at how to prevent safety mechanisms from being altered and redistributed.

At this point, only limited information is available on the specific distribution channels and operators involved, making it urgent to get a clear picture of what is actually out there.

Tailwinds for Open Weights, an Opening for Mistral

With turmoil at major US tech companies dragging on, WIRED argues that "the moment for open-weight AI models to steal the spotlight" has arrived — a development seen as playing into the hands of France's **Mistral**.

The advantages of open weights — models that anyone can freely download and modify — are being reassessed, and the focus is on whether Mistral can set itself apart from the closed-model giants. Competition is fierce, though, and it remains unclear whether Mistral can hold on to an edge.

DeepSeek-V4-Flash: Just How Fast Can It Get Locally?

On the LocalLLAMA forums, tinkering around self-hosting **DeepSeek-V4-Flash** remains as active as ever. It has been the running theme for days now, and new benchmark numbers keep rolling in.

Running on Mac (M3 Ultra)**: Using the MXFP4 MLX-quantized build of DeepSeek-V4-Flash-0731 with DeepSpark/MTP support included, one report says code generation improved from 34 to 43 tok/s — reportedly the fastest result on Mac at the moment.

Optimizing on dual GH200s**: Building vLLM v0.26.0 from source, applying the DSV4 cache layout patch (PR #48993), and setting DeepSpark's speculative tokens to 6 delivered roughly 276 tok/s while handling 1M context on 192GB of HBM. SGLang, meanwhile, was faster across all decode workloads, reaching about 317 tok/s.

The Q8 quantization trap**: Running Q8_K_XL on the Vulkan build of llama.cpp reportedly falls into a so-called "doom loop" repeating "Need maybe," and users are debating whether the latest build is required.

Moving off LM Studio**: Some users are weighing a switch to llama.cpp proper, and the choice of GUI or harness has become a talking point.

All of these are community-measured numbers, so bear in mind that they depend heavily on hardware and settings.

Cutting Chinese Components: A New Direction in AI Robot Sourcing

A push is emerging to pare China out of AI robot supply chains as much as possible. **Ati Robotics**, a startup featured by WIRED, assembles its robots in India and relies on only a small fraction of Chinese-made components.

As the Trump administration tightens the screws on Chinese humanoid robots, this stance could become a real differentiator. At the same time, a complete break from China is difficult, and how far the company can diversify its components will be the key.

Closing Thoughts

Volta attacking the AI chip "capital wall" with financial engineering, the safety risks that open weights bring, and the optimization race in local inference — all of these feel like a cross-section of a moment when "AI democratization" and "safety and control" are being called into question at the same time.