SAP's €2bn Bet on Using Whatever Model Works

The strongest AI strategy this week was not choosing a single best model. SAP, Oracle and OpenAI are spreading work across models and chips, while agent products are adding context, permissions and evaluation. As intelligence becomes easier to swap, the durable advantage shifts to the system around it.

For years, model choice has been treated like a foundational decision: pick the smartest provider, build around it and hope the lead holds. This week made that approach look increasingly temporary. The more interesting companies are assuming that models, chips and even agent components will change, then designing the rest of the system so the business does not have to start again every time they do.

SAP stopped picking winners

SAP says it uses more than 70 AI models from providers including OpenAI, Anthropic and Google, and it has started replacing expensive models where cheaper alternatives are good enough for the task.¹ The company expects this approach to contribute to more than €2 billion in AI-related efficiency gains over the next two years. That is a striking number because the saving does not come from discovering one model that beats everything else. It comes from refusing to treat any model as sacred.

OpenAI appears to be making a similar calculation at the hardware layer. Reuters reported that it plans to buy Google's TPUs after renting Google's chips since 2025, while remaining a major Nvidia customer and developing its own chip with Broadcom.² Oracle is going even further by putting Google TPUs, Nvidia Vera Rubin systems, Cerebras inference and Intel CPUs inside the same cloud portfolio.³The winning architecture increasingly looks like a portfolio of options, not a permanent bet on one supplier.

That matters because AI workloads are splitting into very different jobs. Training a frontier model, running a long-context inference request, executing an agent tool call, classifying a document and generating a short piece of copy do not need the same level of intelligence or the same hardware. Paying for the strongest available model on every task is the computational equivalent of sending a senior lawyer to photocopy a contract. The capability may be excellent, but the economics are absurd.

Most artificial intelligence news still leads with benchmark gains and new model names because they are easy to compare. The more consequential question for buyers is becoming whether a system can route work to the cheapest model that meets the quality threshold, then move again when price, latency or capability changes. That requires good evaluation, clear task definitions and enough engineering discipline to know what "good enough" actually means. A company that cannot measure quality has no safe way to take advantage of cheaper models.

This is where the affordability story gets more interesting for smaller companies. Generative AI gets cheaper when model prices fall, but it becomes genuinely more accessible when products hide that complexity and choose the right level of intelligence for each job. A small business should not need a model procurement strategy to use AI content tools. It should pay for useful work, while the product underneath makes sensible choices about cost and capability.

Agents need explicit boundaries

Model choice becomes less important when AI moves from answering questions to taking actions. Salesforce announced new Agent Evals, API Evals and spending controls for Agent Fabric, including the ability to intervene when quality drops, cap spending and take an agent offline where policy allows.⁴ Salesforce says Agent Fabric handles more than 1.2 billion requests a day. At that volume, a slightly smarter model matters less if the surrounding system cannot detect a bad loop, a broken API call or an agent that has crossed a boundary.

Agent identity is becoming a market in its own right. Mintive announced funding for a product that gives enterprise agents their own identities and authorisations, controlling which services they can reach, which actions they may perform and which models they can use.⁵ The product exists because an agent acting inside business systems cannot safely inherit vague, unlimited access from whoever launched it. Once software can browse, read files or take action, identity and permission design become product features, not administrative chores.

This changes what people should mean when they say an agent is "autonomous". Autonomy is not a binary state where the software either acts alone or waits for a human. Useful autonomy is bounded: the agent has a defined identity, a limited mandate, a known set of tools, spending limits, approval points and an audit trail. The broader the mandate, the more carefully those boundaries need to be designed.

That is also why the human role does not disappear as agents improve. People move from repeating every routine step to deciding what the agent may do, which exceptions matter, when evidence is sufficient and what should happen when the system is uncertain. Those decisions are harder to automate because they depend on consequences, organisational priorities and accountability. More capable software increases the value of well-placed human judgement rather than eliminating the need for it.

Context became part of the product

Several of this week's launches were built around a simple admission: a general model does not know enough about the job. Circuit Mind's AVE verifies circuit-board designs against datasheets, physics constraints and each customer's engineering rules.⁶ Jedify's Data Agents use a context graph containing an organisation's metrics, business rules, relationships and permissions before they investigate what changed.⁷ Legora gives its legal workflow an agreed playbook for NDA review rather than asking a general model to improvise legal policy.⁸

That grounding is what turns a capable model into a useful system. The model can recognise language and patterns, but it does not know what "revenue" means inside a particular company, which component an engineer has approved, or what clause a legal team is willing to accept without escalation. Those definitions often look boring compared with a model demo. They are also the part that customers cannot replace with a cheaper API call.

This is especially relevant to AI content generation for small business. An Instagram AI content tool can generate a caption quickly, but speed is not the same as usefulness if the system does not know the business, its products, its visual assets, its tone and what it would never say. How to automate Instagram content creation is therefore partly a context problem: the product has to keep the business's own material and identity attached to the generation process. Otherwise cheaper generation simply produces more interchangeable content.

The same logic explains why vertical products keep attracting attention. Dream is building models grounded in physics, building standards and permit data for construction, while Casap applies institutional policies and system access to bank disputes. The valuable part is not the generic ability to produce words or recommendations. It is the accumulated context that lets the system act inside a real job without pretending every company works the same way.

More agents can make things worse

The research this week is a useful corrective to the idea that more context and more agents automatically produce better decisions. TeamGym found that when pivotal facts were split across agent participants, success fell by an average of 22.9 percentage points compared with giving the same information to one model.⁹ Requiring agents to retrieve the relevant facts themselves created a further 19.9-point gap. A team can collectively contain the right information and still fail to move it to the participant who needs it.

Another paper, "Rare, Relevant, and Still Wrong", found that compressing a multi-agent discussion into memory can preserve misleading evidence, and that a full discussion transcript could perform worse than giving the final model no notes at all.¹⁰ The researchers also found a social-proof effect: facts attributed to several agents gained extra weight even when that apparent consensus added no evidence. Removing or randomising attribution improved accuracy by 13 to 24 points in the paper's classic test setting. More information can therefore create more confidence without creating more truth.

A separate study of persistent coding agents found that every tested system lost 13 to 23 points under heavy shared CPU and memory workloads compared with isolated tasks.¹¹ The result matters because most benchmark demonstrations happen in unusually clean conditions. Real businesses have queues, retries, concurrency, broken dependencies and competing demands on the same resources. An agent that succeeds beautifully in isolation may behave very differently once five other jobs arrive at the same time.

These findings make evaluation a first-class part of the product. Teams need to know whether failure came from reasoning, retrieval, communication, scheduling, bad memory or a weak acceptance rule. Without that diagnosis, the default response is to buy a stronger model and hope the problem goes away. Sometimes the model was not the problem at all.

What does the model still own?

The answer is still a great deal. Frontier models remain expensive to train, difficult to reproduce and central to what these systems can do. The enormous infrastructure spending around Meta, Microsoft and others shows that raw capability is not becoming irrelevant. But the week's pattern suggests that capability is becoming easier to buy than a reliable operating system around it.

Open standards reinforce that shift. The Agentic AI Foundation announced donated projects from Anthropic, OpenAI, Hugging Face and Alibaba covering agent skills, response formats, environments and sandboxes, while saying it had passed 300 member organisations in less than a year.¹² When common components become easier to connect, a proprietary wrapper around one model becomes less defensible. Domain knowledge, customer context, permissions, evaluation and workflow design become harder to copy because they have to be learned inside the job.

That is a healthy direction for the market. It gives buyers more negotiating power and gives smaller software companies a better chance to build useful products without owning a frontier model. It also puts pressure on vendors whose only advantage is access to a capability that every competitor can buy next month. Better models will keep arriving, but the product still has to know what to do with them.

For workers, the same shift is more hopeful than the usual replacement story. Legora says its workflow returned 9.5 hours per lawyer each week in one reported measure, while other systems this week focused on maintenance, analytics and engineering checks.⁸ The interesting question is what people do with the returned time, and whether the system gives them better evidence for the decisions only they can own. Productivity is valuable when it expands what a person can do, not when it merely produces more machine activity.

Build for the next swap

SAP's 70-model setup is useful because it treats change as normal. The company is not betting that one provider will stay cheapest, fastest or best for every task. It is building around the assumption that the answer will keep changing, then trying to capture the economic benefit each time it does.

That should influence how companies buy and build AI now. A good architecture should make it possible to replace a model without rebuilding the business rules, customer context, permissions and evaluation that sit around it. If switching the model breaks the product, too much of the product was tied to the wrong layer. If switching the model lowers cost while everything important stays intact, the system is working as intended.

The same test applies to agents. Give them enough context to be useful, enough authority to save time and enough limits that a mistake is visible and reversible. Keep people at the points where judgement, accountability and exceptions matter. The companies that do this well will benefit every time models get cheaper or better, because improvement arrives as an input rather than a rewrite.

This week's most useful signal is therefore not which model won a benchmark. It is that SAP can use more than 70, Oracle can offer several kinds of compute, Salesforce can stop an agent, and researchers can show that extra agents sometimes make decisions worse. The competitive advantage is moving into the choices around the model, because those choices determine whether intelligence turns into useful work.

Sources