Four developments from the past two days cut across every layer of the AI stack: the researcher most associated with the attention mechanism that underpins modern language models is switching companies, Amazon is moving toward selling its custom AI chips to anyone who wants them, the US grid regulator has ordered a structural change to how data centres connect to the power network, and a new open-weight model rewrites how text generation can work. For operators, the collective message is that the infrastructure assumptions of twelve months ago are all in motion at once.
Noam Shazeer, Co-Author of "Attention Is All You Need," Leaves Google for OpenAI
On 17 June, Google DeepMind vice president of engineering and Gemini co-lead Noam Shazeer announced he is leaving to join OpenAI as its Lead for AI Architecture Research. Shazeer is a co-author of the 2017 paper "Attention Is All You Need," the work that introduced the Transformer architecture and made modern large language models possible. He is arguably the most consequential individual talent hire the industry has seen.
The context adds weight. Shazeer left Google in 2021 to co-found Character.AI; Google brought him back in 2024 in a deal reported at $2.7 billion. Two years into that return, leading the Gemini programme, he is leaving for the direct competitor. OpenAI chief executive Sam Altman said the partnership was "ten years in the making." Also joining OpenAI this week: Dean Ball, a former Trump White House AI policy official whose addition signals OpenAI's intention to maintain political access ahead of its IPO.
For operators, the hire is a signal rather than an immediate product change. It means OpenAI is investing at the foundational architecture layer -- not only in scaling or distribution -- at the precise moment it is preparing to go public. If Shazeer alters the core design of future OpenAI models, the competitive landscape for reasoning, efficiency, and inference cost could shift materially within one to two model generations.
Amazon in Talks to Sell Trainium3 Chips Directly to Outside Data Centres
For the first time since it designed its own AI accelerators, Amazon is in active discussions to sell Trainium3 chips to external data centre operators, directly competing with Nvidia in the hardware sales business. Amazon's AI infrastructure head Peter DeSantis confirmed the talks on 18 June without naming prospective customers. Bloomberg reported the discussions on the same day.
The strategic shift matters for several reasons. Amazon's chip division already generates an estimated $20 billion in annual revenue from internal AWS workloads. Trainium3 is reported to deliver four times the performance of Trainium2 at roughly half the cost of equivalent Nvidia hardware, and units have run at near-capacity utilisation since their late-2025 launch. DeSantis's case for selling externally: demand for AI compute still exceeds supply across the entire industry, so selling chips to competitors' data centres does not cannibalise cloud revenue.
The near-term practical implication for operators locked into Nvidia pricing is limited -- enterprise procurement cycles are long and the talks are preliminary. The medium-term implication is more significant: Amazon entering the external chip market puts genuine competitive pressure on Nvidia's pricing for the 2027-2028 procurement cycle, which is when many enterprises will be renegotiating AI compute contracts at scale.
FERC Orders US Grid Operators to Create a Fast Lane for AI Data Centres
On 18 June, the Federal Energy Regulatory Commission issued orders under Section 206 of the Federal Power Act to all six major US regional grid operators, directing each to develop expedited interconnection pathways for large electricity consumers drawing more than 20 megawatts. The orders follow a formal 2025 request by Department of Energy Secretary Chris Wright and represent the first time a federal regulator has issued tailored mandates to every US regional grid operator simultaneously.
The specific mechanisms include allowing data centres to co-locate with generation assets, compressing interconnection study timelines, and shifting a greater share of interconnection costs directly onto the requesting entity rather than distributing them across the grid. The last point raises upfront capital requirements for a new facility but clears the queue of speculative applications that have made typical US interconnection timelines stretch to four or more years.
For teams planning AI infrastructure in the United States, the orders do not eliminate grid delay -- study timelines remain measured in months, not weeks -- but they change the calculus for data centre siting. Co-location with generation assets is no longer a niche workaround; it is now a formally supported pathway backed by federal regulatory authority. Operators planning facilities in the two-to-four year horizon should model this option explicitly.
Google DeepMind Releases DiffusionGemma 26B: A Faster Open Architecture with a Real Trade-Off
On 10 June, Google DeepMind published DiffusionGemma 26B-A4B under an Apache 2.0 licence on Hugging Face, Kaggle, and Vertex AI. The model is a 26-billion-parameter mixture-of-experts, with approximately 3.8 billion active parameters per pass, that abandons autoregressive left-to-right generation in favour of a diffusion process: it initialises a block of output with placeholder tokens and iteratively refines them in parallel, drawing on the same principles used in image diffusion models.
The stated speed advantage is substantial. Google cites up to four times faster generation than equivalent autoregressive models, reaching over 1,100 tokens per second on a single NVIDIA H100. The honest caveat from Google's own documentation is equally notable: the company explicitly recommends standard Gemma 4 "for applications that demand maximum quality." DiffusionGemma trades benchmark quality for throughput.
The operator case is specific. Workloads that are latency-sensitive, quality-tolerant, and expensive to run at autoregressive speed -- real-time summarisation at scale, structured output for downstream processing, draft generation pipelines -- are worth evaluating on this architecture. For any workload where accuracy is the primary constraint, this release changes nothing today. The longer-range signal is architectural: if diffusion-based text generation closes the quality gap as the technique matures, the entire inference stack built around autoregressive models becomes a candidate for redesign.
The through-line across all four items: the AI stack is being rebuilt simultaneously at the talent, silicon, infrastructure, and architecture layers. That rarely happens in a single week. Operators who have been deferring decisions about compute strategy, data centre siting, and model architecture evaluation now have concrete reasons to bring those conversations forward.