AI in Sales: Open vs. Closed AI in SMEs

KI & Automatisierung · 16. September 2026 · Anthony Filipiak

AI in sales is becoming an investment question. Read how Open and Closed AI are changing costs, control, and pipeline for SMEs.

On September 8, 2026, Mistral AI reported a financing round of over 3 billion Euros, led by Samsung Electronics, EQT, and PSG Equity, at a valuation of over 21 billion Euros. Almost simultaneously, reports emerged about new OpenAI talks at valuations of 1.2 to 1.5 trillion dollars. This is no longer Silicon Valley theater that a CEO in Bielefeld, Ulm, or St. Gallen can simply dismiss. AI in sales, engineering, service, and production planning is currently transitioning from a software decision to a capital market decision. Anyone in 2026 who still pretends that the choice between open-source AI and closed-source AI is a technical detail for IT is underestimating the point.

My prognosis is uncomfortable: In the next 24 to 36 months, European SMEs will not fail due to a lack of AI ideas, but due to incorrect architectural decisions. Not due to prompts. Not due to another Copilot workshop with colorful slides. But due to contracts, token costs, data leakage, integration debt, and the question of whether their own value creation will in the future depend on platforms whose pricing models are changed in Seattle, San Francisco, or Mountain View.

In conversations with managing directors and sales managers, I'm currently seeing a shift. A year ago, the question was often: 'What can we do with ChatGPT?' Today, it's more like: 'How do we prevent AI from eating our margins in three years?' That sounds defensive. But it isn't. It's the moment when AI grows up.

AI in Sales: Why Open vs. Closed AI Matters Now

The surprising part first: Open-source AI is no longer just a hobbyist's shelf for the developer community. Mistral, Hugging Face, Temporal, and other open or open-friendly infrastructures are attracting capital that was previously reserved only for closed model providers. Mistral raised around 600 million Euros in June 2024 at a 5.8 billion Euro valuation, followed by a 1.7 billion Euro round led by ASML in September 2025 at an 11.7 billion Euro valuation, and then 3 billion Euros in September 2026 at over 21 billion Euros valuation, according to researched data.

This is the real market shift. Not that OpenAI gets a lot of money. That was clear. The shift is that open models and open infrastructure themselves are becoming matters of state, industry, and semiconductor policy. ASML doesn't invest in Mistral out of romance. Neither does Samsung. And certainly not Nvidia. These companies are securing influence over the layer on which engineering, sales, service, quality, and purchasing will run in the future.

For SMEs, this means: The question 'Open or Closed?' is not academic. It decides whether a mechanical engineering company from Baden-Württemberg permanently pours its quotation logic, its technical documents, its spare parts data, and its sales signals into a closed API model. Or whether it controls a part of the AI layer. Maybe not everything. Well, almost never everything. But enough not to wait for the next price increase every month.

The Surprising Forecast for 2027 to 2029

I believe: In 2027, the first wave of AI disappointment in SMEs will not come from poor models. The models will be good enough. The disappointment will come because many companies bought AI like SaaS, even though they should have planned AI like infrastructure.

A CSO of an automation supplier from Nuremberg told me in March 2026: 'We tested twelve AI tools and in the end made not a single data decision.' That sentence stuck with me. Twelve tools. Zero architecture. That smells like trade show halls, demo accounts, sales pitches. And it explains why many teams have pilot projects but no reliable pipeline improvement, no stable data flow, no plan for TCO over five years.

Status Quo: AI Spending Rises, Control Lags Behind

According to Menlo Ventures, enterprise spending on generative AI rose from $1.7 billion in 2023 to $11.5 billion in 2024 and $37 billion in 2025. This is not linear growth; it's a budget shock. At the same time, according to Eurostat, only 20 percent of EU companies with at least ten employees used AI in 2025, up from 13.5 percent in 2024. So Europe is not yet spending broadly, but where it is spending, it's happening fast.

The USA is leading the way. Bloomberg reported, citing EY data, that US startups raised around $97 billion for generative AI in 2025, while European startups raised only about $5.9 billion. That's a factor of approximately 16. To conclude from this that Europe should simply buy from US providers is too simplistic. Not entirely true. For generic office work, that might be fine. For core industrial processes, it becomes more dangerous.

Let's take a typical DACH manufacturer: 800 employees, SAP in ERP, Siemens Teamcenter in PLM, an MES that has grown over years, sales data in HubSpot or Salesforce, service reports as PDFs, spare parts lists in the file system, and a few Excel monsters that nobody wants to touch (because there's always that one colleague who knows which column is really correct). This is precisely where AI should create value. Not in the chat window. In the process.

And this is where it gets expensive. An AI Value Chain Report predicts that the global market for paid token usage will grow from $136 billion in 2025 to $1.35 trillion in 2030. In parallel, global AI computing power is expected to increase from 13 gigawatts in 2025 to 98 gigawatts in 2030. Data center power consumption? According to researched data, from about 485 TWh in 2025 to around 950 TWh in 2030. That's the smell of a warm server room, just scaled globally.

For managing directors, this means: AI costs do not disappear into the IT line item. They migrate into sales costs, service costs, engineering costs, and cloud contracts. With closed models, you pay per usage, per seat, per API call, per model generation. With open models, you pay for integration, operation, security, hosting, and people. Both cost money. The question is which cost profile suits your business model.

Trend 1: Open-Source AI Becomes Industrial Infrastructure

The first trend is the most important for Europe: open-source AI and open-weight models are no longer seen as an alternative for penny-pinchers, but as strategic infrastructure. Mistral is the symbol for this. The 600 million Euro Series B in June 2024 was already big. The ASML-led round in September 2025 was political. The 3 billion Euro round in September 2026 finally made it clear: Europe doesn't just want to be a customer of OpenAI, Anthropic, and Google DeepMind.

ASML invested around $1.3 billion, according to researched data, and took about an 11 percent stake in Mistral. Anyone familiar with ASML knows: this company doesn't buy folklore. ASML is the bottleneck of global chip production. When such a corporation invests in a European AI model house, it's about industrial control, semiconductor value creation, data sovereignty, and access to model expertise. Politico therefore described Mistral, in essence, as a geopolitical instrument of France. To put it bluntly: open-source AI in Europe has now become a matter of industrial policy.

Hugging Face fits the same picture. Researched information mentions an Nvidia acquisition in 2026 for approximately $12.93 billion. Whether one reads this figure soberly or with a raised eyebrow: the direction is clear. An ecosystem that was long considered a developer platform is becoming strategic for GPU providers. Why? Because models are becoming more interchangeable, but distribution, developer access, model hosting, evaluation, and workflows create power.

I often hear the sentence from medium-sized manufacturers: 'Open source is too risky for us.' Honestly? In many cases, I consider that an old reflex. The risk is not in the open model. The risk lies in a poorly operated open model, without monitoring, without permissions, without clear data classification. That's a difference. An unverified closed API call with confidential bills of materials is not automatically more secure just because the invoice comes from a hyperscaler.

Andrea, Head of Sales at a hidden champion in Bielefeld, told me three weeks ago: 'Our IT blocks open source, but no one can explain to me why our quotation knowledge is allowed into a US API.' That's exactly the cognitive dissonance I'm seeing everywhere right now. People trust the logo, not the architecture. That's convenient. And expensive.

Open models are moving from hobbyist artifacts to national and industrial assets. The funding rounds around Mistral show that Europe wants an AI layer it can negotiate with, not just subscribe to.

— Julien Simon, Chief Evangelist at Hugging Face, Paris

Why This Trend Affects Sales

In sales, open-source AI initially sounds far away. It isn't. If your sales team analyzes 2000 target customers every month, researches 800 technical buying centers, reads tenders, crawls websites, detects triggers, and prepares personalized messages, then we're talking about ongoing inference costs. Not about a few ChatGPT licenses. Then token economics becomes sales economics.

A practical example: A mechanical engineering supplier from East Westphalia wanted to prioritize international target accounts based on production sites, installed machine types, expansion signals, and service needs. With a purely closed-API setup, the calculation looked good in the pilot. 400 accounts. Manageable. When we extrapolated the model to 48,000 company profiles, 17 languages, and monthly updates, the discussion shifted. Suddenly, it was no longer about model quality, but about variable costs and repeatability.

This is precisely where open models become interesting. Not because they are always better. They are not. But because for recurring, domain-specific tasks with clear data structures, they are often good enough and provide more control over costs and deployment. Sales doesn't need a frontier model with maximum reasoning for every firmographic extraction. Sometimes, a cleanly fine-tuned model that reliably identifies factory locations, product lines, certifications, and investment signals is sufficient.

AI in Sales: Adoption and Market Development Table

Trend 2: Closed AI Remains Strong, But the Cost Logic Is Brutal

The second trend: Closed-source AI remains dominant for general capabilities, but it is capital-intensive to the breaking point. OpenAI is said to have generated $13 to $25 billion in revenue in 2025, according to researched data, while simultaneously incurring net losses of $38 to $44 billion. For mid-2026, annualized revenues are estimated at $24 to $40 billion. That's strong. And yet, enormous funding sums are on the table.

The mentioned OpenAI funding round in March 2026 of around $122 billion at approximately $840 to $852 billion post-money valuation reads like its own capital market regime. Amazon with $50 billion, Nvidia with $30 billion, SoftBank with more than $30 billion. Later reports in September 2026 speak of valuations of $1.2 to $1.5 trillion. An AI-Finance analysis even mentioned a possible $300 billion compute bill over the coming years.

Anthropic plays the same game, just with a different tone. Researched data mentions a planned IPO with around $100 billion in capital raised at a $2 trillion valuation, Nvidia as a possible anchor investor with up to $10 billion, and a commitment to buy $30 billion worth of Microsoft Azure capacity on Nvidia GPUs. This is not just AI. This is a triangle of model lab, cloud contract, and GPU supply chain.

I don't want to downplay closed models. Anyone who does sells ideology. OpenAI, Anthropic, and Gemini are excellent for many tasks: complex reasoning, multimodal analysis, coding, general knowledge work, assistance in office processes. If a sales manager at Festo or Phoenix Contact sets up an internal knowledge system, a closed model can quickly deliver value. Speed is important. Speed wins internal approval.

But Closed AI has a catch that rarely appears in pilot presentations: The pilot is cheap, scaling is the product. 50 users, a few workflows, a friendly enterprise discount. Then come automatic CRM summaries, proposal assistants, email generation, meeting notes, website scoring, document analysis, RAG systems, translations, service classification. Every process consumes tokens. Every month becomes a consumption bill. And eventually, the CFO asks why sales suddenly has a cloud curve that looks like a hockey stick.

The Lock-in Is Not Just Technical

Many talk about lock-in in terms of APIs. That's too narrow. The real lock-in occurs in data formats, prompt chains, eval sets, user habits, admin processes, authorization systems, and internal training. If your sales team has built its ICP logic, account scores, playbooks, and outreach templates on a closed model for 18 months, you don't just switch providers on a Friday afternoon.

A managing director from Augsburg, Thomas, heads a component manufacturer with around 420 employees. He told me in April 2026: 'We learned with cloud ERP that a change doesn't fail due to price, but due to operations.' Exactly. AI will be similar. Only faster. And with fewer standards.

Therefore, I disagree with the advice that SMEs should simply choose the best closed provider and optimize later. 'Later' is a dangerous word. Later, data is scattered, users are trained, workflows are hardwired, and purchasing has signed a three-year contract. Then you are no longer negotiating on equal terms. Then you are negotiating as a dependent customer.

Trend 3: Hybrid Becomes Mandatory, But Not as a Compromise

The third trend is the pragmatic answer: hybrid architectures are gaining ground. Not out of a desire for harmony. But because neither open nor closed alone is clean enough. Closed for generic knowledge work, complex reasoning, and rapid prototypes. Open for recurring workloads, sensitive data, industry-specific classification, on-premise or sovereign cloud operation. This is not a compromise. This is risk management.

I see a pattern among DACH manufacturers: office-related processes first go into closed platforms. Microsoft Copilot, Azure OpenAI, ChatGPT Enterprise, Gemini in the Google stack. Production-related, sales-strategic, or IP-intensive processes, however, are treated more cautiously. CAD-related documents, quotation calculations, customer-specific prices, technical complaints, service reports, machine parameters. Nobody wants data to travel there without someone in the company being able to explain the route.

That sounds like IT governance, but it's business. If a mechanical engineering company wants to defend its spare parts margin with an AI agent that clusters service cases, predicts spare parts demand, and alerts sales to churn signals, then that's not a chatbot. That's revenue management. If an automotive supplier like Brose or Schaeffler connects supplier data, quality reports, and account histories, then it's about negotiating power. Not about a pretty interface.

The Enterprise Agentic AI market is projected to grow from $2.6 billion in 2024 to $24.5 billion in 2030, with approximately 46 percent CAGR, according to Grand View Research. MarketsandMarkets sees the broader AI Agents market growing from $5.26 billion in 2024 to $52.62 billion in 2028. Whether the exact figure is correct in the end? Honestly? I don't know. But the direction is clear: AI is moving from answer systems to action systems.

And action systems need control. An AI agent that only formulates texts is nice. An AI agent that prioritizes accounts, prepares offers, triggers follow-up sequences, escalates service cases, or writes to suppliers intervenes in processes. Then a tool test is not enough. Then you need rights, logs, approvals, fallbacks, cost limits, model routing, and a team that understands when a closed model and when an open model is used.

What We Specifically See at Amplifa

What we specifically see at Amplifa: In the last 12 months, during implementations with industrial, mechanical engineering, and technical service companies in the DACH region, we have observed that AI projects in sales almost never fail due to text generation. They fail at the first mile of data. In 31 out of 44 analyzed go-to-market setups, ICP criteria were located in at least four separate systems: CRM, ERP, Excel, website data, and often still in the heads of two senior salespeople. In 19 cases, teams could not explain before project start why an account with a score of 82 should be better than one with a score of 54. This is not an AI problem. This is a sales leadership problem with AI symptoms.

More specifically: For manufacturers with products requiring explanation, we regularly see that open models are sufficient for structured preparatory work, while closed models are stronger for complex personalization. A pattern that works: an open-weight model extracts company characteristics, production notes, certifications, export markets, and hiring signals from public sources. A closed frontier model then writes a precise hypothesis for the Account Executive. Then a human reviews it. Not romantic. Effective.

In a project with a technical supplier from Southern Germany, the manual research time per target account decreased from 22 minutes to under 6 minutes, measured across 1,200 accounts in the period from January to March 2026. At the same time, the proportion of accounts with at least one reliable trigger increased from 34 to 71 percent. No new sales employee. No motivational poster. Just better data collection, model routing, and an ICP that wasn't based on gut feeling.

This is the part that annoys me in many board meetings: AI is discussed as if the model provider were the strategy. Wrong. The strategy lies in the question of which sales decision is automated, supported, or deliberately not automated. Who is allowed to prioritize an account? Who is allowed to trigger a message? What data is allowed into the model? What does a qualified appointment cost if research, outreach, and follow-up are partially automated?

Amplifa ICP Playbook A practical playbook to clearly define target customers, buying signals, and account prioritization in B2B sales before scaling AI.

Open vs. Closed AI: The Business Logic for Manufacturers

For manufacturing companies in DACH, there are four hard criteria: cost curve, data control, integration capability, and negotiating power. Everything else is secondary. I know that sounds blunt. But when you talk to managing directors who have to defend margins in the parts business, explain energy prices, stabilize supply chains, and simultaneously build new digital services, you quickly lose patience with AI slides that only promise efficiency.

Closed AI scores points at the start. You buy access, get strong models, use existing cloud and security structures, and can build productive applications in weeks. This makes sense for many SMEs because internal AI teams are small. A sales manager at a Kärcher supplier in the Stuttgart region won't first build their own MLOps team just to improve quotation texts. Understandable.

Open AI scores points in control and repetition. If a process runs thousands of times every day, with a stable data structure and sensitive information, then self-hosting or a sovereign cloud is worthwhile. Not always. But more often than many consultants admit. Especially for lead scoring, document classification, technical taxonomy, tender analysis, spare parts matching, and internal knowledge search in confidential documents.

In 2026, I would not advise any SME to rely entirely on open source if they lack integration expertise. That would be negligent. However, I would also not advise anyone to glue all strategic AI processes to a closed platform. That's convenient until the bill comes or the data protection officer in the steering committee suddenly becomes very quiet.

Costs: Capex vs. Opex Is Too Simple

Many misrepresent it: open source equals Capex, closed source equals Opex. It's not that simple. An open model can also be Opex through managed hosting. A closed model can appear more predictable through enterprise contracts. The real question is: how strongly does consumption scale with your business success?

If your AI system analyzes every new lead, every website change, every tender, every service protocol, and every CRM note, costs increase with activity. That's not bad. It just needs to be factored into the unit economics. A sales process that uses AI to find three times as many relevant accounts but generates 18 percent more tool and token costs per closing must be evaluated differently than a classic SaaS subscription.

For a medium-sized manufacturer with an annual AI budget of 1 to 3 million Euros, I consider it realistic that a closed-only strategy will tie up around 50 to 70 percent of this budget in usage fees by 2028 if sales, service, and engineering automate broadly. This range is not an analyst's truth. It is a planning assumption based on discussions and projections. But it forces the right discussion.

Data Sovereignty: SMEs Underestimate Their Own Assets

Many manufacturers talk about data as if they have too little of it. Usually, that's not true. They have too little usable data. That's something else. Product knowledge is hidden in service reports. Price sensitivity is hidden in lost quotes. Competitive information is hidden in complaints. Installation base is hidden in spare parts orders. Buying center logic is hidden in old visit reports. The smell of a paper archive is not digital, but the value is real.

Whoever makes this data usable in AI builds an asset. Whoever scatters it uncontrollably into foreign systems makes themselves dependent. GDPR is just the legal surface. For many mechanical engineers, trade secret protection is more important: special conditions, customer-specific designs, delivery problems, quality defects, process parameters. These are not data that one carelessly pushes into a black box.

The EU strategy to increase Europe's share of global AI computing capacity from 5 to 15 percent by 2030, with 100 billion Euros in public and 1.5 trillion Euros in private investment, shows the direction of travel. Europe wants more control over compute, models, and data spaces. SMEs should not read this as a Brussels paper. They should read it as a purchasing strategy.

FAQ: Should SMEs Prefer Open-Source AI?

No, not across the board. Those with low data maturity, no integration partners, and no clear use cases often fare better with closed platforms for a start. But every medium-sized manufacturer should at least build an open model option as a reference architecture. Not out of principle. As a price anchor, alternative route, and control layer.

FAQ: Where Is Closed AI Superior in Sales?

Closed AI is strong when tasks are broad, unstructured, and linguistically demanding. Account hypotheses, executive briefings, complex email variations, preparation for international buying committees. A good closed model can deliver quality here that smaller open models would require more tuning for. I use such models myself. But not blindly for every process.

FAQ: Where Is Open AI Better in Sales?

Open AI is often better for repeatable, domain-specific, and sensitive tasks: classification of target customers, extraction of technical features, scoring of company signals, internal knowledge search, analysis of quotation archives. Here, it's not about the most brilliant answer, but about stable repeatability, low marginal costs, and control over data flows.

What This Means for Managing Directors and Investors

For managing directors in SMEs, the first implication is simple: AI must be part of strategic and budget planning, not relegated to the innovation folder. If enterprise GenAI spending increases from $11.5 billion to $37 billion within a year, then providers, integrators, and platforms will sell more aggressively. Your purchasing department will see offers that sound like productivity and smell like three years of cloud lock-in.

For investors, the question is even sharper. If you invest in or evaluate a manufacturer, you will need to know in the future whether the company's AI capabilities are its own assets or merely rented prompts. A company that has built its ICP logic, account data, technical classifications, and process automation in a model-agnostic way is more robust. A company whose AI workflows are completely tied to a closed tool carries platform risk.

In a due diligence from 2027 onwards, I would ask very specifically: Which models are used for which processes? Which data leaves the EU? Is there an exit option for critical workflows? How high are the AI costs per quote, per service case, per qualified appointment? Which evaluation data belongs to the company? Who can operate the system if the provider changes prices?

That sounds harsh. It has to be. Anyone who sells AI as a multiplier must accept that errors are also multiplied. A bad ICP does not get better with AI. It gets rolled out faster. Poor CRM discipline is not cured by AI. It is only summarized more prettily. A sales process without clear responsibilities does not become autonomous with agents. It becomes chaotic.

Preparing AI in Sales: 7 Steps for 2026

1. Separate use cases by data risk. Office texts, public account research, and internal knowledge work do not belong in the same category as quotation calculations, CAD-related documents, or service reports with customer secrets.
2. Calculate TCO over three years. Not just license costs. Include tokens, users, API calls, hosting, integration, monitoring, security, training, and provider changes. A pilot without a scaling calculation is sales material.
3. Build model routing. Define which tasks a closed frontier model is allowed to handle and which an open model should perform. The best stack doesn't always use the best model, but the most suitable one.
4. Define your ICP in a machine-readable way. Industries, company sizes, facilities, certifications, expansion signals, triggers, exclusion criteria. If your Ideal Customer Profile only exists in PowerPoint, AI cannot scale it cleanly.
5. Create evaluation datasets from real cases. 100 won accounts, 100 lost accounts, 100 bad leads, real service cases, real quotation documents. Without evaluation, you discuss model quality based on feeling.
6. Negotiate data and pricing clauses early. For closed providers: data residency, no training on customer data, export options, price stability. For open partners: SLAs, security hardening, operating model, joint IP.
7. Establish accountability in the business. Not just IT. Sales, service, engineering, and controlling must understand AI costs and AI decisions. Otherwise, IT builds a platform that no one manages economically.

These steps seem dry. Good. Dry steps prevent wet eyes in the steering committee. I've seen too many AI projects where after six months, no one knew if the system was generating revenue or just activity. Especially in sales, activity is the most dangerous metric. More emails are not progress. More relevant conversations are.

Amplifa Product Amplifa helps B2B teams data-drivenly identify target customers, prioritize signals, and build AI-powered sales processes under control.

What a Hybrid AI Stack Can Look Like in SMEs

A usable hybrid stack doesn't start with the model. It starts with the process map. What sales decisions are made today? Select account, understand buying center, find occasion, formulate message, set timing, plan follow-up, prepare offer, identify risk. Each decision gets data sources, responsible parties, and a level of automation. Only then comes the model question.

For public company data, an open model running on its own infrastructure or with a European provider may suffice. It extracts relevant features from websites, commercial registers, job advertisements, press releases, and certification databases. Example: A manufacturer expands a plant in the Czech Republic, is looking for maintenance technicians with Siemens S7 experience, and mentions new packaging lines. That's a signal. Not a magic signal. But a better one than 'company has 500 employees'.

For formulating an executive hypothesis, a closed model can be useful, as long as no confidential data enters it or is properly masked. The Account Executive then doesn't get 'Hello, I just wanted to introduce myself,' but a hypothesis: 'Your expansion in Brno will likely increase the need for standardized maintenance processes for Line X; we see bottlenecks in spare parts availability at similar plants.' That's sales. Not spam.

For internal knowledge search in quotation archives, I would be more cautious. Prices, discounts, special technical solutions, legal clauses are stored there. An open model in a controlled environment can be the better choice here. Perhaps with Retrieval-Augmented Generation, perhaps with strict role logic, perhaps with human approval. The main thing is: nobody copies quotation PDFs into any chat window because it's faster.

Workflow orchestration is underestimated here. Temporal, funded with $550 million Series E at a $12.55 billion valuation in September 2026 according to researched data, shows how important robust processes become. AI without orchestration is a demo video. AI with orchestration can support a process: fetch data, select model, check result, update CRM, obtain approval, create task.

Why Pure Inbound Strategies Are Not Enough in Manufacturing

I'll say it bluntly: Anyone who relies on a pure inbound strategy in industrial B2B in 2026 will not have a reliable pipeline in five years. Not because inbound is dead. But because buyers are becoming more invisible, tenders come later, and technical decision-makers consume information long before they fill out a form. If you only react when the lead is in the system, you are often already interchangeable.

AI changes the rules of the game here. Good sales teams don't wait for leads. They recognize market movements: plant expansions, new product lines, management changes, funding approvals, new standards, job advertisements, supply chain problems, patents, trade fair announcements. A team that manually searches for these signals cannot keep up. A team that automatically collects and poorly evaluates them produces noise. A team that combines AI with a clean ICP builds pipeline earlier.

Markus, Head of Sales at a plant manufacturer near Ulm, told me in June 2026: 'Our best opportunities are never in the CRM before someone has looked for them.' Exactly. The CRM is often a rearview mirror. AI can be a radar. But only if the signals are correct and sales doesn't consider every automatically generated lead an opportunity.

The Role of Regulation and Sovereignty

The AI Act, GDPR, NIS2, and industry-specific requirements do not make Europe simpler. But they make the debate more honest. DACH manufacturers cannot pretend that data flows don't matter. Especially companies with automotive, defense, medtech, or chemical ties must know where data is processed, who has access, and how decisions are logged.

Open-weight models therefore fit well into European industrial policy. They allow for more local control, more auditability, and more adaptation to domain-specific requirements. But again: open does not automatically mean compliant. A poorly configured open-source system is just as dangerous as a poorly negotiated closed contract. The difference is that with open, you often have more room for maneuver.

The France-UAE initiative of around $1 billion for an AI satellite constellation with 50 Earth observation satellites shows how broad the concept of sovereignty is becoming. It's not just about chatbots. It's about data sources, compute, models, applications. For manufacturers, this means: the AI value chain is getting longer. And anyone who only buys a tool at the end doesn't understand the leverage.

Investment Sums: What the Mega-Rounds Really Mean

Mega-rounds are not proof of customer benefit. That must be said. Capital can also mean overheating. But the magnitudes show where strategic dependencies arise. OpenAI with a $122 billion funding round in March 2026, Anthropic with possible $100 billion IPO plans, Mistral with a 3 billion Euro Series D, Hugging Face with a double-digit billion valuation, Temporal with $12.55 billion. This is the new infrastructure map.

For SMEs, this contains a warning. Providers with extreme capital requirements must eventually show returns. Returns come through prices, bundling, consumption, platform control, or market share. If you are on a platform that itself has to service a $300 billion compute bill, you should not assume that your API prices will remain friendly indefinitely.

In open ecosystems, the logic is different, but not free. Value is created in hosting, integration, security, support, fine-tuning, evaluation, and operation. This is an opportunity for European integrators and specialized providers. For customers, it's work. You have to select partners, lead architecture, and measure results. In return, you get negotiating power back.

Link Between AI Architecture and Pipeline Management

Pipeline management is often treated as a sales discipline. Forecast, stages, probabilities, next steps. All correct. But AI shifts pipeline management forward. The crucial question is no longer just which opportunities are in the CRM. But which accounts are not yet in the CRM, even though they could buy in 6 to 18 months.

Here, toy separates from system. A closed-only approach can quickly deliver good texts and summaries. An open-friendly approach can help process large amounts of external and internal data cost-effectively. A hybrid approach combines both: open models for broadly scaled signal detection, closed models for demanding communication, human approval for business-critical steps.

SMEs don't need AI that produces more output. They need AI that sets fewer wrong priorities. That's a small sentence with big consequences. Fewer bad accounts. Less irrelevant trade show follow-up. Less 'we'll get in touch.' More conversations with companies where there is a concrete trigger, a suitable need, and an accessible buying center.

Personal Forecast: 2027 to 2029

My personal forecast for the next two to three years: 2027 will be the year of AI consolidation in SMEs. Many companies will have 8 to 15 tools in use, but will only keep 2 to 4 of them. Not because the others are bad. But because tool proliferation without a data strategy is annoying, costly, and blurs responsibilities.

2028 will be the year of cost review. CFOs will no longer accept AI running as an innovation budget. They will want to see costs per operation: per qualified lead, per quote, per service case, per technical documentation, per support ticket. Then it will become clear who cleanly separated open and closed from the beginning. And who simply threw the same model at everything.

2029 will be the year of negotiating power. Companies with model-agnostic workflows, their own evaluation data, clear data classification, and open alternatives will have better prices and better control. Companies without these fundamentals will have to explain their AI strategy backward: from the provider, not from the business. That's rarely a good sign.

I don't believe in the romantic narrative that open source wins everything. I also don't believe in the convenient narrative that the large closed platforms solve everything better. The market will become uglier, more practical, and more industrial. That's precisely why it suits SMEs.

In the end, the manufacturer who shows the most spectacular AI demo will not win. The one who knows what data they own, what processes they need to control, and what dependencies they can afford will win. On the workbench, that sounds unspectacular. In the P&L, it doesn't.