Here are the five AI developments that matter most today, and the bigger implications behind them.
1. Anthropic's agent market found a more basic problem than negotiation
Anthropic put Claude agents into a controlled market where 201 employees across six offices asked agents to trade books on their behalf. After a short conversation about reading preferences, the agents' rankings matched participants' own rankings on 61% of book pairs. The agents then negotiated with one another on a digital trading floor.
The surprising result was that bargaining was not the main weakness. The market scored 0.55 on Anthropic's efficiency measure versus a theoretical optimum of 0.89, and Anthropic attributes 85% of that gap to imperfectly representing what people actually wanted. The experiment was internal, low-stakes and limited to books, so it should not be generalized too far.
The big picture
Agentic commerce may depend less on making agents tougher negotiators and more on giving them an accurate, permissioned understanding of the person they represent.
That makes preference data, memory, identity and context strategically important. An agent that can bargain brilliantly but misunderstands the customer will still make the wrong decision efficiently.
2. Anthropic's $11.6 billion Akamai deal says AI demand is spreading beyond GPUs
Akamai announced a seven-year, $11.6 billion cloud commitment from Anthropic, with the potential to expand by another $9 billion. The striking detail is the workload: Akamai says the capacity is intended to support Anthropic's accelerating CPU demand, not just GPU-heavy model training. Akamai expects roughly $5.5 billion in capital spending related to the initial commitment.
TechCrunch notes that the agreement depends on Akamai meeting delivery and service-availability requirements and can be terminated under certain conditions. Akamai's future revenue and capacity estimates are company projections, not guaranteed outcomes.
The big picture
AI infrastructure is becoming more heterogeneous. Training frontier models still requires enormous accelerator capacity, but agents also browse, run code, call tools, move data and orchestrate conventional software. Those activities create demand for large amounts of general-purpose compute.
The infrastructure story is broadening from who has the GPUs to who can supply the right mix of compute, memory, networking and geographic distribution for always-on agent workloads.
3. Google is testing a retailer-branded checkout inside Gemini
Google is testing direct purchases from Walmart-owned Flipkart inside Gemini and AI Mode for some users in India. TechCrunch reports that select product listings now show a Buy button that opens a Flipkart-branded checkout flow without taking the shopper out of the AI interface. The test is limited to certain users and products, and Google has not publicly committed to a broad rollout.
Google has separately confirmed that Flipkart is partnering on its agentic commerce work in India using the Universal Commerce Protocol, which is designed to let agents interact with merchant inventories and checkout systems.
The big picture
The important design choice is that the AI can become the discovery surface without necessarily erasing the retailer from the transaction.
That points toward a more negotiated version of agentic commerce. AI platforms may own the conversation while merchants retain pieces of checkout, branding, fulfillment and customer economics. The battle is not only over who gets the sale. It is over which layer owns the customer relationship and the data created along the way.
Read more: TechCrunch, Google tests Flipkart purchases inside Gemini and AI Mode | Google, agentic commerce in India
4. Oracle's Stargate project is running into the physical limits of AI
Oracle sent a force majeure notice to the developer of Project Jupiter, a 2.45-gigawatt Stargate data center campus planned for New Mexico. Oracle says it still expects the project to stay on schedule and is not trying to exit the deal. The notice would protect Oracle if the facility misses its targeted 2028 opening.
The risk is largely physical. TechCrunch reports that a gas pipeline intended to serve the site has been delayed nearly six months after permit denials, while a separate air-quality permit for the fuel-cell system is still pending.
The big picture
AI strategy now reaches all the way down to pipelines, substations, permitting and local politics.
Capital and chips do not guarantee capacity. As data-center projects get larger, power availability and regulatory execution become part of the product roadmap. The companies planning the biggest AI deployments increasingly have to understand infrastructure risk with the same seriousness they apply to model risk.
Read more: TechCrunch, Oracle sends force majeure notice on Project Jupiter | Axios, Oracle moves to shield itself from AI buildout risks
5. NetApp is buying for the next AI bottleneck: moving and organizing data
NetApp announced plans to acquire U.K.-based PEAK:AIO, a specialist in metadata architecture and high-performance parallel file systems. The company says the technology is designed to let metadata scale independently from stored data and support shared AI storage across very large GPU clusters. No purchase price or closing date was disclosed, and the transaction remains subject to customary approvals.
NetApp says the planned architecture is intended to support trillions of files and multi-exabyte environments. Those are design targets, not reported production results.
The big picture
Compute gets most of the attention, but AI systems also have to find, move, protect and serve enormous amounts of data fast enough to keep expensive processors busy.
As AI infrastructure matures, storage architecture and metadata become strategic bottlenecks. The question is not only how much compute an enterprise can buy. It is whether its data layer can feed that compute without creating a new constraint.
Read more: NetApp, intent to acquire PEAK:AIO
THE THROUGH LINE
The model is becoming only one component of the AI system.
Market agents need an accurate understanding of the people they represent. Cloud providers need new mixes of CPU, memory and distributed infrastructure. Commerce platforms need rules for who owns the interface and transaction. Data centers need power and permits. Enterprise AI needs storage systems capable of moving and organizing data at a very different scale.
The next phase of AI will be decided increasingly by the infrastructure around the intelligence: context, compute, distribution, data and control.
That is a less glamorous story than another benchmark jump, but it is where much of the durable value, and much of the risk, is moving.