Compute & Infrastructure

Nvidia is in advanced negotiations to acquire Hugging Face at a valuation of approximately $14 billion, according to Bloomberg. The deal has not been confirmed and Bloomberg's sourcing is described as people familiar with the matter, so it remains in the category of advanced but unannounced. Hugging Face is the dominant platform for open-source model hosting, fine-tuning tooling, and ML developer workflows — with tens of thousands of models hosted and a developer community that spans every major AI lab and enterprise user.

For Nvidia, ownership of Hugging Face would extend its competitive moat well beyond chip silicon. The company already controls the de facto standard for GPU compute (CUDA), and Hugging Face controls the dominant distribution layer for AI models. Combined, Nvidia would influence where models are trained, how they are optimised, and which hardware they are deployed on. This has significant implications for cloud providers — AWS, Azure, and Google Cloud — who depend on Hugging Face as a neutral third-party ecosystem. Regulatory scrutiny, particularly from the EU's Digital Markets Act apparatus and U.S. antitrust bodies, should be considered a near-certainty if the deal closes.

Nvidia has invested $3.5 billion into Taiwanese fabless designer MediaTek, with the companies agreeing to integrate NVLink Fusion into MediaTek's custom XPU design platform, according to Tom's Hardware and ServeTheHome. The partnership also encompasses local AI computing and automotive platforms. This is a confirmed investment deal.

NVLink Fusion allows third-party CPUs and custom accelerators to connect to Nvidia's GPU ecosystem via Nvidia's high-bandwidth interconnect, rather than PCIe. By embedding NVLink Fusion into MediaTek's custom chip offerings, Nvidia gains architectural influence over a new class of AI infrastructure buyers — hyperscalers and national compute programmes that want custom silicon but still need Nvidia-compatible interconnect. Separately, Nvidia's Hot Chips 2026 presentation on RISC-V integration into CUDA and NVLink Fusion systems, noted by ServeTheHome, signals that Nvidia is actively designing its ecosystem to accommodate open-ISA compute cores, reducing customer friction in adopting heterogeneous architectures that still depend on Nvidia's interconnect fabric.

At SEMICON Taiwan 2026, Phison Electronics CEO KS Pua stated that NAND flash memory is on course for its worst-ever shortage in 2027, driven in part by AI infrastructure demand, according to Bloomberg. Phison designs controller chips for NAND used in SSDs and storage arrays — placing the company directly in the signal path for enterprise storage procurement. This is an executive-level forecast, not confirmed supply data, but Pua's visibility into the controller supply chain gives it operational credibility.

SEMI CEO Ajit Manocha's concurrent commentary on memory prices and an ongoing semiconductor supercycle, also from SEMICON Taiwan, reinforces the directional signal. GPTC's CEO discussion of surging demand for advanced packaging adds another layer: the constraint is not just wafer capacity but back-end packaging, where CoWoS and other advanced techniques remain bottlenecked at TSMC. For data centre operators, a NAND shortage in 2027 would raise storage costs precisely as AI inference workloads — which require large, fast local storage for model weights and KV-cache — scale aggressively.

GMI Cloud, an Nvidia partner operating in Asia, secured commitments exceeding double its target in a loan backed by contracted GPU compute revenue, according to Bloomberg. Separately, Lambda Labs reportedly secured $1 billion in private debt to purchase Nvidia GPUs, with those chips to be leased to Microsoft, per Data Center Dynamics. Anthropic has reportedly signed a $35 billion cloud agreement with Lambda for a data centre in Nueces County, Texas, per Data Center Dynamics. The Lambda-Anthropic figure is reported without confirmed contract documentation and should be treated as unverified.

The structure emerging here — private debt backed by GPU assets or contracted compute revenue, with the chips leased to hyperscalers or frontier AI labs — represents a maturation of AI infrastructure finance. It mirrors the project finance structures used in renewable energy, where physical assets with contracted cash flows support large debt tranches. The oversubscription of GMI Cloud's loan indicates institutional capital is pricing AI compute demand as a low-default-risk contracted revenue stream, at least at current utilisation rates.

OnZero has partnered with Finnish utility Helen to connect a Helsinki AI data centre to the city's district heating network, with capacity to supply up to 525,000 MWh of heat annually, according to Data Center Dynamics. This is a confirmed partnership, though operational commissioning timelines are not specified. Nordic cities with district heating infrastructure are uniquely positioned to monetise data centre waste heat, and this arrangement converts a cost centre — cooling — into a partial revenue offset, improving the economics of northern European AI compute locations.

BT's plan to strip 200,000 tonnes of copper from its legacy network — with potential proceeds of $2.7 billion driven by AI-era copper demand for data centre power distribution — illustrates a secondary materials dynamic worth tracking, per Tom's Hardware. Data centres are intensive copper consumers for busbars, power cabling, and liquid cooling manifolds, and the AI buildout is visibly moving commodity pricing in base metals.