Avalerion Intelligence · Edition 012

Tuesday 25 August 2026 · AI, Data & Enterprise Transformation for the Nordics, UK & Ireland

Nvidia warns AI servers cost 15 percent more

This week at a glance

1. Top AI News

Nvidia told its biggest customers that the hardware gets dearer, and blamed memory.

What happened. Bloomberg reported on 22 August that Nvidia had notified major customers of price increases, in many cases more than 15 percent, on servers built around its accelerators. The increases are reported to apply to systems delivered from the beginning of 2027, across both the Grace Blackwell generation and the coming Vera Rubin platform. Nvidia has not published or confirmed these figures, so treat the number as a well sourced agency report rather than a price list. The stated cause is memory. TrendForce expects server DRAM contract prices to rise 13 to 18 percent in the third quarter of 2026, with supply constrained into 2027, and memory now accounts for roughly a quarter of what a high end rack costs to build.

Why it matters. This is a different animal from the model price moves of recent months. An API price is a commercial decision a vendor can reverse. A component shortage is a physical constraint that reaches every quote in the chain, including the hosted service you buy instead of hardware. Your cloud provider is buying these systems.

Business implications. Any 2027 budget built on 2026 unit economics is understating infrastructure by a double digit percentage, whether or not you ever buy a server.

What to do. Ask every AI infrastructure supplier one question before you sign: is this price fixed for the term, or indexed to component cost. The answer changes the number you should put in the plan.

Google gave away control of the protocol its competitors depend on.

What happened. Google transferred the Agent2Agent protocol to the Agentic AI Foundation, a Linux Foundation body. Axios reported the move on 17 August and the foundation confirmed it on 20 August. A2A now sits alongside the Model Context Protocol, donated by Anthropic, under the same neutral governance. The two remain separate projects with their own technical steering committees. The foundation has grown from 49 founding members to more than 250 in under a year, with AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI at the top tier.

Why it matters. A2A is how agents from different vendors talk to each other. MCP is how an agent reaches your tools and data. With both under neutral stewardship, the two interfaces that decide whether your agents are portable are no longer any single vendor's property.

Business implications. Interoperability moved from a promise in a roadmap to a specification you can name in a contract, which is a far stronger position for a buyer.

What to do. Put both protocols in your next agent related RFP by name, and ask for the vendor's conformance position in writing.

Nvidia paid 6 billion dollars for a toolchain rather than building one.

What happened. Nvidia agreed to pay Poolside about 6 billion dollars to license its Model Factory software, plus roughly 1 billion invested at a 12 billion dollar pre money valuation, with offers extended to more than 100 staff. The licence is reported as non exclusive, leaving Poolside free to license elsewhere. The deal was broken by Newcomer and reported around 20 to 21 August. Nvidia has not issued a detailed public confirmation.

Why it matters. The company with the most compute and the deepest engineering bench decided that buying the machinery for building models beat building it. That is a signal about where durable advantage now sits, and it is not in the toolchain.

Business implications. If Nvidia will not build its own model tooling, your organisation almost certainly should not either.

What to do. Review any internal platform work that duplicates a mature commercial or open component, and be honest about what it actually differentiates.

2. Enterprise AI Trend: the switching cost moved into the contract

For two years, the honest answer to whether you could change agent vendors later was no, because the integration was the product. That is changing. With A2A and MCP under one neutral foundation and 250 members signed up, the technical barrier to moving between agent platforms is falling towards the level of the commercial barrier. That is a meaningful shift, because commercial barriers are things a procurement team can negotiate and technical ones are not. The practical consequence is that lock in is becoming visible. It now lives in your contract terms, your data egress arrangements and your custom code, rather than hiding in a protocol nobody else implements. Visible lock in is manageable lock in. The organisations that benefit are the ones that write portability into agreements now, while vendors are competing to look open, rather than in two years when a renewal is on the table and the leverage has gone.

3. Data & AI Readiness: someone has to check the output

The quiet signal this week came from a small round. OrbDB, a Stockholm company, raised roughly 15.8 million Swedish kronor in pre seed funding to build infrastructure for verifying AI outputs. That a verification layer is now a fundable category tells you something about where deployments are failing. Most enterprise AI programmes have a model, a data pipeline and a dashboard. Very few have a defensible answer to the question of who checks that the output was right, on what sample, against what ground truth, and what happens when it was not. Evaluation is treated as a launch activity rather than an operating one. That gap is precisely what turns a working pilot into a system nobody will accept accountability for in production. If your AI programme has no named owner for output verification and no scheduled evaluation cadence, that is the readiness gap to close this quarter, ahead of any new model or platform.

4. Business Process Spotlight: a component shortage is not a negotiation

Most procurement processes are built to handle vendor pricing. A supplier raises a price, you benchmark alternatives, you negotiate, you switch if you must. None of that works on a component shortage. When memory is the constraint and memory is a quarter of the cost of the rack, every alternative supplier is facing the same input price, so competitive tension does not deliver relief. This is a category of cost risk many finance functions have not modelled since the semiconductor shortages of the early 2020s, and it has now reappeared inside the AI budget line. The right response is not to hunt for a cheaper vendor. It is to change the shape of the commitment: understand which of your workloads genuinely need frontier hardware, fix pricing where you can get a term, and make sure someone has actually calculated what a 15 percent infrastructure increase does to the business case you approved. Two of the three answers are architecture decisions.

5. Nordic Technology: Sweden is funding the unglamorous layer

Swedish rounds in the week of 17 August ran to eleven disclosed deals, and the pattern is more interesting than the total. SweGaN in Linköping raised 14 million dollars in Series B for gallium nitride semiconductor materials, which is a bet on the physical layer at the exact moment component supply is deciding AI economics. OrbDB is building output verification. Solva raised 6 million dollars in seed for AI applied to insurance outcomes. Plock, a usage based billing platform, was acquired by Kleer, an AI finance and payroll platform, which matters more than the deal size suggests: as vendors move to metered AI pricing, the ability to bill and be billed by consumption becomes infrastructure. Mentimeter reported annual recurring revenue of 614 million kronor and a return to profitability. This is not the headline generating end of the market. It is components, verification, billing and profitable software, which is roughly the list of things an AI estate needs and rarely has.

Do this next

Three questions this week put on the table, thirty minutes with us: whether your 2027 AI infrastructure budget assumes today's component prices or next year's, whether your agent contracts name the interoperability standards your vendors have now publicly committed to, and who in your organisation owns verification of AI output once a pilot goes live. We will map all three with you. No pitch, no deck.

Book an AI Readiness Session

Recommended reading: What AI readiness actually means for Nordic companies and Why most enterprise AI projects still fail, at avalerions.com/insights

Frequently asked questions

Sources