OpenAI will train GPT-5 on your data. Is your data ready?
Avalerion Intelligence · Edition 004
Tuesday 14 July 2026 · AI, Data & Enterprise Transformation for the Nordics, UK & Ireland
This week at a glance
| Move | Who | Why it matters to you |
|---|---|---|
| GPT-5 fine-tuning opens to enterprises while self-serve winds down | OpenAI | Custom models become a premium tier, and your data quality is the ceiling |
| Hard API cutover on 24 July: deepseek-chat and deepseek-reasoner start returning errors | DeepSeek | If your code calls those model names, it breaks in ten days |
| Beijing weighs restricting overseas access to top Chinese models, open weights included | Reuters, 7 July | The cheap-model strategy carries a geopolitical dependency |
| Up to 8,900 forward-deployed AI engineers, plus acquisitions | TCS | Even the firms that sell engineering hours are rebuilding around deployment |
| Record 26.5 billion dollar raise for AI memory | SK Hynix | The compute bill keeps compounding, and it arrives in your prices |
| Narvik campus gets an operating company and a long-term power deal | Nscale, Nordkraft, Vattenfall | Europe's AI compute is moving to where the power is, which is here |
1. Top AI News
Custom AI becomes a large-enterprise product
What happened: OpenAI opened early access to fine-tuning for GPT-5, letting enterprise customers train private versions of the model on their own data, terminology and knowledge domains rather than assembling retrieval pipelines around a generic model. The pricing puts it firmly out of reach of individual developers and small teams. In the same period, OpenAI has been winding down its self-serve fine-tuning platform: new organisations are already blocked from creating fine-tuning jobs, and new job creation ends for existing customers by January 2027. Why it matters: the industry spent three years insisting the model was the product. This week the most valuable AI company started selling something else, which is your own knowledge, reflected back at you. Business implications: if a competitor fine-tunes on fifteen years of well-structured claims decisions and you fine-tune on a shared drive, you both bought the same model and only one of you bought an advantage. Fine-tuning does not create knowledge. It compresses whatever you already have, including the mess. What to do: before you price a fine-tuning engagement, answer one question honestly. Could you hand a supplier 10,000 examples of your best work, correctly labelled, in a consistent format, with the wrong answers excluded? If not, that is the project. The fine-tune is the easy part.
There is a hard deadline in your codebase, and it is ten days away
What happened: DeepSeek's hosted API completes a forced migration on 24 July at 15:59 UTC. After that, calls to deepseek-chat and deepseek-reasoner return errors; teams must point at deepseek-v4-pro or deepseek-v4-flash instead. Separately, Reuters reported on 7 July that Chinese authorities have met Alibaba, ByteDance and Z.ai about restricting overseas access to China's most advanced models, potentially including open-weight releases and models not yet published. Talks are exploratory and DeepSeek was reportedly not in the room. Why it matters: many Western teams quietly routed workloads to cheap Chinese models over the past year. The saving was real. The dependency came with it, and this month it produced both a breaking change and a political question. Business implications: a one-line model-name change is trivial if you know where the calls are. It is a fire drill if you do not. And a supplier whose government may restrict export is a supplier you cannot plan a three-year roadmap around. What to do: today, search your codebase and your vendors' integrations for deepseek-chat and deepseek-reasoner and fix them before 24 July. This week, write down every external model your organisation calls, who calls it, and what happens on the day it stops answering.
The physical bill keeps compounding
What happened: SK Hynix raised a record 26.5 billion dollars to expand AI memory production as demand for high-bandwidth memory outruns supply. In parallel, the US Energy Information Administration projects national power consumption climbing from 4,195 billion kWh in 2025 to 4,399 billion in 2027, with AI data centres a principal driver. Why it matters: the marginal cost of intelligence is falling while the total cost of the infrastructure behind it is exploding, and someone pays for both. Business implications: today's token prices reflect a capacity land grab, not a settled market. Build your business case on the value of the outcome, not on the assumption that inference stays cheap forever. What to do: stress-test your two largest AI workloads at triple today's token price. If the case dies, it was never a case; it was a subsidy.
2. Enterprise AI Trend: even the integrators are rebuilding themselves
What happened: TCS, India's largest IT services firm, will convert 1% to 1.5% of its associate base into forward-deployed AI engineers, roughly 5,900 to 8,900 people from a headcount of 593,798, and is hunting acquisitions in AI and cybersecurity after years of purely organic growth. CEO K Krithivasan describes the role plainly: engineers who sit inside the client's business and make the model actually do something. Why it matters: the company whose entire business model is selling engineering hours has concluded that selling engineering hours is no longer enough. Business implications: the industry has now converged on one answer to why AI does not land, and it is not the model. It is that nobody restructured the work around it. Expect every supplier you have to arrive with an embedded-engineer offer this autumn. What to do: decide now what you want from those people. A vendor engineer who maps your process and hands you the map leaves you stronger. One who builds a black box around their employer's stack leaves you rentable. Ask for the map in the contract.
3. Data & AI Readiness: what "fine-tunable" actually means
What happened: fine-tuning has quietly become the clearest audit of data maturity anyone has invented, because it fails loudly and cheaply. A model trained on inconsistent examples produces inconsistent output, immediately and visibly. Why it matters: most organisations discover the true state of their data only at the moment they try to use it for something that cannot tolerate mess. Business implications: the readiness gap is not abstract. It is the difference between having 10,000 clean, labelled examples of a decision your business makes well, and having a folder. What to do: pick one high-value decision your organisation repeats: a credit approval, a claims assessment, a quote, a support resolution. Then check four things. Are the past decisions recorded in one place? Is the outcome of each one recorded next to it? Can you separate the good ones from the bad ones? And does anyone own that data set? Four yeses and you are fine-tunable. Anything less and you have found this quarter's real project.
4. Business Process Spotlight: find your gold standard
What happened: every fine-tune needs a gold standard, which is a set of examples that represent the work done right. Almost no organisation has one written down, which is why so many AI projects begin by inventing one under deadline pressure. Why it matters: if you cannot say what "done right" looks like for a process, no model can learn it and no agent can be trusted with it. Business implications: the gold standard is not an AI artefact. It is a management artefact you should have had anyway, and the firms that already have it are the ones whose AI projects finish. What to do: take your highest-volume repeated decision and ask the person who does it best to work through twenty real cases, writing down why each answer is the right one. That document is your training data, your quality benchmark, your onboarding material and your agent's specification. It costs two days. It is the cheapest asset you will build this year.
5. Nordic Technology: the compute moves to where the power is
What happened: Nscale and Nordkraft formed a joint operating company, Nordscale Operations AS, to run the AI data centre at Kvandal near Narvik, which is being built out to 230 MW and will host more than 30,000 Nvidia Rubin GPUs for Microsoft from 2027. Vattenfall signed a long-term renewable power purchase agreement covering a large share of the site's electricity from 2027 to 2031. The think tank Ember expects Nordic data centre energy demand to grow four to five times over the next decade, putting Norway and Sweden alongside Europe's largest hubs. Why it matters: while the rest of Europe debates where AI power will come from, the Nordics are signing the contracts. Business implications: the region's advantage is no longer just clean, cheap electricity. It is becoming a location advantage for anyone who needs European data residency and a compute bill that does not move with a scarcity market. What to do: the next time you renew a cloud or AI contract, ask two specific questions: where does this run, and what does the power cost. If the answer is a Nordic site, you have a defensible position on cost, carbon and residency. If nobody at the table knows, that is your answer.
Avalerion's Take
They will sell you the model. They cannot sell you the memory. This week the pitch changed. OpenAI will now train GPT-5 on your data. TCS will send you thousands of engineers to make it work. SK Hynix is raising twenty-six billion dollars to keep the chips coming. Every one of those offers assumes something the vendors cannot supply: that your organisation knows what it does, has written it down, and can hand it over. That is the whole asset. A model fine-tuned on excellent, labelled, owned data is a genuine advantage, and the only one on this list that a competitor cannot buy with a purchase order. A model fine-tuned on a shared drive is an expensive mirror. The uncomfortable part is that the work is not technical and never was: define the decision, capture the outcome, separate good from bad, give it an owner. Nobody will do that for you, because nobody else can. Architecture before acceleration.
Could you hand a supplier 10,000 clean, labelled examples of your best work tomorrow? For most organisations the answer is no, and that answer decides what any AI investment is worth. 30 minutes, no pitch, no deck.
Book an AI Readiness SessionFrequently asked questions
What is GPT-5 fine-tuning and who can use it?
OpenAI opened early access to fine-tuning for GPT-5 in July 2026, letting enterprise customers train private versions of the model on proprietary data, terminology and knowledge. The pricing targets large enterprises. At the same time OpenAI is winding down self-serve fine-tuning: new organisations are blocked from creating jobs, and new job creation ends for all customers by January 2027.
Is fine-tuning better than RAG for enterprise AI?
They solve different problems. Retrieval gives a model access to documents at query time and is easier to update. Fine-tuning bakes tone, format, terminology and decision patterns into the model itself. Fine-tuning only pays off when you have a large, clean, consistently labelled set of examples of the work done correctly. Without that, retrieval is usually the better first step.
What happens to the DeepSeek API on 24 July 2026?
DeepSeek completes a forced migration at 15:59 UTC on 24 July 2026. After that, calls to deepseek-chat and deepseek-reasoner return errors. Teams must switch the model parameter to deepseek-v4-pro or deepseek-v4-flash, using the same base URL and API key. It is a one-line change per call, provided you know where the calls are.
Could China restrict access to its open-source AI models?
Reuters reported on 7 July 2026 that Chinese authorities met Alibaba, ByteDance and Z.ai to discuss restricting overseas access to the country's most advanced models, potentially including open-weight releases and unreleased systems. The talks are exploratory with no confirmed outcome, but any organisation depending on Chinese models should have a documented fallback.
Why are AI data centres being built in the Nordics?
Abundant low-cost renewable power, a cool climate and fast grid connections. Nscale and Nordkraft launched an operating company for the 230 MW Kvandal campus near Narvik, which will host over 30,000 Nvidia Rubin GPUs for Microsoft from 2027, backed by a long-term Vattenfall renewable power agreement. Ember expects Nordic data centre energy demand to grow four to five times over the next decade.
Avalerion Intelligence is the weekly briefing of Avalerion Consulting AB on AI, Data & Enterprise Transformation for the Nordics, UK and Ireland. Subscribe by email. Published by Avalerion Consulting AB, Malmö and Stockholm.
