AI Infrastructure: Foxconn, Bull, and AI Factories

KI & Automatisierung · 3. Juni 2026 · Joseph Flesh

European AI infrastructure is transforming smart factories and sales. Read what steps SMEs in DACH should now plan for AI Factories.

Last Thursday, 8:12 AM, I was on an onboarding call at Amplifa with Andrea, COO of an automation company from Ulm. In the background, I heard the faint screech of a saw from production; she had only half-closed her office door, and on her desk lay a printout of delivery dates, marked in red. “Joseph, we have leads for three new lines,” she said, “but our customers are now asking about AI infrastructure first, and only then about the mechanics.” I briefly looked at my coffee. Cold. Then the news about the new Foxconn-Bull partnership appeared on my second screen.

My prognosis is simple and uncomfortable: Any company in mechanical engineering, electronics manufacturing, or automation components that doesn't have a clear answer for AI Factories by 2028 will not first lose orders in production. They will lose them in sales. Not because every 180-person company will suddenly operate its own GPU clusters. Well, almost. But because customers in DACH are starting to sort suppliers based on whether they can truly deliver data, Edge AI, digital twins, and plant integration — or if they just write “Industry 4.0” on slides.

AI Infrastructure in Europe: Why This News Should Hurt

Foxconn, officially Hon Hai Technology Group, and Bull SAS, the French IT infrastructure subsidiary from the Atos environment, plan to jointly build AI servers, GPU clusters, edge systems, and automation hardware in Europe. At first glance, this sounds like a data center. Like sheet metal, racks, cables, fans. Like something that COOs are interested in and sales managers skip over. That's precisely where the misconception lies.

According to industry reports published in the last 24 to 48 hours, the partnership explicitly targets European manufacturing companies, Smart Factories, and so-called AI Factories. Germany, Austria, and Switzerland are named as core markets. Automotive. Electronics. Process industry. Precision manufacturing. In other words, exactly the companies we regularly work with at Amplifa: 50 to 500 employees, a good product, a customer base grown over years, but a pipeline that increasingly depends on projects where IT, OT, and purchasing all have a say.

For me, this news is not a hardware detail. It's a sales signal. When Foxconn and Bull push “AI infrastructure made in Europe” into the market, customer expectations for Trumpf suppliers, Festo integrators, Schaeffler partners, or Webasto-related component manufacturers shift. The customer no longer asks: “Can you deliver the plant?” They ask: “Can you integrate it into our data architecture, extend it with computer vision, evaluate maintenance data, and operate the whole thing in an auditable way?” That's a different sales process. Tougher. More technical. Longer. But also more lucrative.

Status Quo: Many Talk About AI, Few Have Laid the Wires

According to the Tacton “State of Manufacturing” Report 2026, 79 percent of manufacturers worldwide are actively investing in or experimenting with AI. The previous year, it was 64 percent. This is a leap that cannot be dismissed as trade fair hype. At the same time, according to the same source, only 7 percent of manufacturers have true end-to-end connectivity across the value chain, sales, engineering, production, and service. I repeat the number because it usually causes a brief silence in workshops: seven percent.

I see this gap every month. In March 2026, I spoke with Markus, sales manager of a special machine builder from Heilbronn. His team sells systems in the range of 250,000 to 1.8 million euros. The company uses a CRM, an ERP, a PLM, Excel quote templates, and three SharePoint folders, all named “final.” The machine itself generates sensor data every second. But in sales, no one knows which features are actually used by existing customers. No one. After-sales has the information, engineering suspects something, but sales still calls with generic arguments.

This is the status quo in many medium-sized manufacturing companies in DACH. There are machines with Siemens controls, camera systems from Cognex or Keyence, robot cells from Kuka or ABB, ERP data from SAP Business One or Microsoft Dynamics, and then there's sales, still prioritizing by gut feeling which OEM is currently “hot.” Not entirely true. The good teams have lists. But lists are not intelligence. And they become worthless as soon as customers start using AI-readiness as a purchasing criterion.

What we specifically see at Amplifa: In 17 implementations at industrial companies between 70 and 430 employees over the last 12 months, we almost always found the same pattern. The top 20 existing customers generate between 52 and 78 percent of the contribution margin, but less than a third of sales activities are based on installed base, machine age, service cases, or spare part patterns. Sales works at the company level, but the opportunity arises at the plant level. This is precisely where AI Factories become relevant — not as a buzzword, but as a data structure for revenue.

Trend 1: AI Factories Move from the Cloud to the Factory Floor

The first trend is physical. Loud. Warm. You can hear it. Anyone who has stood next to a GPU rack knows that AI doesn't live in a cloudy metaphor, but in power, cooling, latency, cable management, and fire protection. Foxconn and Bull are addressing exactly this: AI servers, GPU clusters, industrial edge gateways, and control hardware are to be manufactured, integrated, and qualified for industrial workloads in Europe.

Why is this relevant for a company with 180 employees in Baden-Württemberg? Because many production-related AI applications don't work cleanly if they only exist as a SaaS demo in the browser. Quality inspection via computer vision requires cameras close to the line, inference times in the millisecond range, and a robust connection to the control system. Predictive maintenance requires machine history, sensor technology, context from maintenance orders, and models that don't fail with every network disruption. Digital twins need more than a pretty 3D model. They need data flows that don't die after the pilot project.

In April 2026, I spoke with Thomas, Head of Sales at an automation integrator in Nuremberg, about a quote for a visual inspection system. His customer, a Tier 2 supplier to the automotive industry, wanted to know if the image data could be processed locally. Not “sometime.” It was in the specifications. The works council had a say, as did IT security. Thomas told me: “We used to sell on cycle time. Now we sell on data residency.” That's exactly the shift.

Siemens Digital Industries argues in a recent article on European AI sovereignty that AI leadership depends not only on models but on semiconductors, compute infrastructure, energy efficiency, robotics, and the ability to translate innovation into production. This sounds like corporate strategy. For SMEs, however, it's a question of offerings. Can I offer the customer a solution that runs in their factory, with their data, under European compliance conditions, and that can still be maintained after six months?

Europe's AI sovereignty will not only be decided by software, but by the ability to bring semiconductors, compute, automation, and industrial data into scalable production.

— Key statement from Siemens Digital Industries, article on AI Sovereignty in Europe, 2026 (paraphrased)

I like this quote because it puts the usual AI debate into perspective. Many sales organizations talk about chatbots, email automation, and prompt libraries. All useful. But in industrial B2B, AI isn't decided in the prompt. It's decided by whether a customer in Linz can expand their line without pushing production data into an unclear cloud region; whether a factory in Bielefeld proactively dispatches spare parts; whether a quality manager in St. Gallen can sleep at night because the model not only had 93 percent accuracy in PowerPoint, but still works in shift three with oil mist and fluctuating light.

Trend 2: Sales Sells Architecture, Not Just Machines Anymore

The second trend directly affects sales managers. Previously, technical sales in mechanical engineering was complex enough: specifications, cycle time, material flow, CE, service, delivery time, price. Now another layer is added. Customers want to know how a solution fits into their AI infrastructure. What data is generated? Where is it stored? What interfaces exist? How are models trained or monitored? Who is liable if the AI makes wrong quality decisions?

I'm blunt about this: Anyone who still believes in 2026 that pure product sales are enough in mechanical engineering has not understood the change. The customer is not just buying an aggregate. They are buying a piece of future operational logic. If your offer doesn't explain this operational logic, someone else will. Perhaps an integrator. Perhaps a large automation company. Perhaps an IT service provider who knows less about spindles than you, but talks better about data flows.

Three weeks ago, I had an internal meeting with Lena, our Customer Success Lead at Amplifa, and we reviewed the reasons for loss from eight industrial pipelines. The room smelled of whiteboard marker because someone left the pen uncapped. In three out of eight cases, the reason for loss in the CRM was “price.” When we reconstructed the email threads and call notes, the real reason was different: The competitor was able to explain earlier how their solution fits into MES, ERP, and local edge architecture. Price was the word sales used because it hurts less.

This is where the Foxconn-Bull announcement becomes interesting. If European AI servers and edge systems become more available, the excuse “It's too early” diminishes. Customers will expect reference architectures. They will ask if your system interacts with local GPU racks, industrial gateways, and standardized data models. And yes, many SMEs will say: “Our customers aren't asking that yet.” Maybe that's true. Maybe they're only asking your competitors.

How Buying Centers Change with AI Infrastructure

In classic mechanical engineering deals, management, production management, purchasing, and perhaps a maintenance manager sat at the table. For AI-Factory-related projects, new voices are added: IT security, data protection, OT managers, sometimes the works council, occasionally an external digitalization consultant. This lengthens deals. It doesn't automatically make them worse. It makes poor sales processes visible.

An example: For a customer in the packaging machinery sector, 220 employees in the Ravensburg region, we rebuilt the ICP logic in January 2026. Previously, sales prioritized by industry and revenue size. Afterwards, criteria such as installed Siemens or Beckhoff controls, number of production sites, service intensity, existing MES, and keywords from job advertisements like “Data Engineer,” “Computer Vision,” or “OT Security” were added. Result after nine weeks: fewer initial conversations, but the conversion rate from discovery to qualified project increased from 31 to 46 percent. No magic. Better targeting.

This is the bridge between AI infrastructure and sales. Hardware availability in Europe creates new project classes. New project classes create new buying criteria. New buying criteria break old lead scoring models. Anyone who still segments only by employee count, postal code, and industry overlooks the companies that are currently freeing up budget for Edge AI, digital twins, and production data platforms.

Trend 3: AI Factories Make Data Sovereignty a Sales Argument

The third trend is political, but not abstract. Data sovereignty becomes commercial. In Germany, Austria, and Switzerland, there is deep skepticism about production data in foreign hands. One can smile about it if one comes from the pure SaaS world. I would not do that. A managing director of a precision parts manufacturer from Villingen-Schwenningen told me in May 2026: “Our process data is our margin.” That statement sticks.

When Foxconn and Bull emphasize European manufacturing, European integration, and local AI infrastructure, they hit precisely this nerve. It's not just about supply chains. It's about trust in tenders. A Swiss pharmaceutical supplier, an Austrian machine builder, or a German automotive supplier must be able to explain to customers, auditors, and internal committees where data is processed, who has access, how updates run, and how systems continue to operate during geopolitical tensions.

According to reports, Bull brings European customer base, integration expertise, and data center design. Foxconn brings manufacturing volume, electronics design, and supply chain power. This is not romantic European self-sufficiency. Foxconn remains Foxconn. But local assembly, qualification, and system integration in the EU can make a difference for industrial customers when delivery times, export controls, or support chains become critical.

For sales, this means: Sovereignty becomes an argument, but only if it becomes concrete. “Made in Europe” alone doesn't sell a plant. “Your image data does not leave the factory, inference runs on an edge cluster near the control cabinet, updates are installed via an approved maintenance window, and the model version is documented in the audit log” — that sells. Or at least it prevents the IT manager from cutting off the conversation after 42 minutes.

FAQ: What is an AI Factory in Industrial SMEs?

An AI Factory is not a single factory with robots and blinking dashboards. In an industrial context, the term refers to a repeatable infrastructure with which companies train, operate, monitor, and integrate AI models into production processes. This includes compute — often GPU servers or specialized edge systems — data pipelines, model management, interfaces to machine controls, cybersecurity, monitoring, and responsibilities. For a 300-person machine builder, this doesn't have to look like Nvidia or BMW. It can be a local edge cluster for quality inspection, supplemented by a digital twin, service data, and a clean MLOps concept. Size is not important. What is important is that AI moves from pilot to operational capability.

AI Infrastructure and Pipeline Management: The Uncomfortable Connection

Many managing directors still too cleanly separate production technology and sales technology. Over there, manufacturing with OPC UA, PLC, MES, OEE, and sensor technology. Here, sales with CRM, campaigns, trade fairs, and visit reports. This separation is organizationally convenient. It is economically dangerous.

If your customers are investing in AI Factories, sales signals emerge long before the official inquiry. A company hires an OT security manager. A plant manager talks about computer vision on LinkedIn. An annual report mentions “local inference.” A site builds a new logistics center. Purchasing suddenly asks about data formats. A designer downloads whitepapers on digital twins. Each signal alone is weak. Together, they form an intent.

This is precisely where many classic CRM setups fail. The CRM stores what sales already knows. It rarely recognizes what's happening outside. At Amplifa, we therefore build ICP and account models that combine external triggers, CRM history, product fit, and timing. Not to replace sales employees. That's nonsense. But to tell them where a conversation makes sense now and where they are just politely disturbing.

Amplifa ICP Playbook A practical playbook to prioritize target customers in industrial B2B based on real buying and technology signals instead of just industry and revenue.

What This Means for SMEs

For manufacturing companies with 50 to 500 employees in DACH, the Foxconn-Bull partnership first means: The bar is rising. Large providers will integrate European AI infrastructure as a building block into their offerings. System integrators will put together complete packages. Customers will get used to AI projects no longer being sold as experiments, but as scalable architecture. Anyone who then only responds with “We can do AI too” will sound like a fax machine in a fiber optic office.

Secondly, margins are shifting. Pure hardware or pure mechanics will continue to come under pressure, especially if Asian and Eastern European competitors remain aggressive on price. Value is created where machines, data, service, and process knowledge are combined. A test bench manufacturer can become a provider of quality intelligence. A component supplier can model failure risks. A special machine builder can create new revenue streams with digital commissioning, simulation data, and service forecasts. But only if sales and engineering speak the same language.

Thirdly, internationalization becomes more selective. Germany, Austria, and Switzerland are prioritized markets according to the partnership, because industrial density, automation expertise, and willingness to pay come together here. That's good. But it also means: More providers will attack precisely these markets. An SME from East Westphalia no longer competes only with the known competitor from the neighboring district, but with European platform partnerships that bundle hardware, integration, and the AI narrative.

I often see a understandable reaction from COOs: wait and see. The line is running, order books are not empty, people are scarce. But waiting is not a neutral position when buying criteria are being rewritten. It is a decision that others set the standards. And standards are brutal in sales. Whoever defines the standard has less to explain. Whoever comes later has to prove.

An Everyday Example: From Spare Part to AI Opportunity

One of our customers from Northern Bavaria sells plant components to manufacturers of packaging and process machinery. Not a corporation, 140 employees, good reputation. In the CRM, existing customers looked the same for a long time: revenue, contact person, last order. When we added service cases, spare part cycles, and publicly visible investment signals, accounts suddenly stood out where old components were running in lines, while the customer was simultaneously advertising positions for “Manufacturing Data Analyst” and “Automation Engineer.” Sales didn't go into the call with “Do you need spare parts?” They went in with “We see in comparable lines that retrofitting and data acquisition pay off together if you want to evaluate quality data locally.” Different tone. Different appointment.

After four months, the team had booked five technical workshops, three of them with production management and IT together. Previously, IT was almost never in talks with these accounts. The point is not that Amplifa had some magic trick. The point is that AI-Factory signals are visible in the market if you look for them. Most sales processes just don't look.

Preparation: 7 Steps for Sales Managers, COOs, and Managing Directors

1. Map your AI-Factory relevance per product line. For each machine, component, or service, write down what data is generated, what interfaces exist, which AI use cases are realistic and which are not. Be honest. A bad AI use case eats trust faster than a delayed delivery date.
2. Expand your ICP with technical triggers. Employee count and industry are not enough. Capture MES projects, OT security roles, new plants, job advertisements for computer vision, ERP migrations, sustainability reports, maintenance initiatives, and hints of digital twins.
3. Build a simple architectural slide that a sales representative can explain. Not 38 boxes. One page: machine, edge, data flow, local processing, cloud option, security, service. If sales doesn't understand this slide, the customer won't either.
4. Train discovery questions for AI infrastructure. Don't ask: “Are you interested in AI?” Ask: “Which production data is allowed to leave your factory?”, “Who is responsible for model approvals?”, “Where do quality decisions fail today?”, “Which line causes the most expensive unplanned downtimes?”
5. Define partner roles early. Not every SME needs to build GPU clusters or operate MLOps themselves. But you need to know whether Bull, Siemens, Phoenix Contact, local system integrators, or cloud partners can be part of your reference architecture.
6. Link service and sales data. Machine age, spare part patterns, fault causes, and maintenance windows are strong buying signals. If this data remains in service while sales makes cold calls, you're burning existing context.
7. Make compliance sellable. EU AI Act, NIS2, Cyber Resilience Act, and data residency should not only appear during contract review. Translate requirements into clear statements that sales and engineering can jointly support.

The fifth point is important to me. Many SMEs are afraid of being technologically left behind in AI infrastructure because they don't have the budgets of DMG Mori, Kärcher, or Brose. This fear is partly justified. But it often leads to the wrong consequence: doing nothing at all because you can't do everything. A clearer focus is better. Which AI function makes your product more valuable to customers? What data do you need for it? What infrastructure must exist at the customer's site or with partners? What statement can your sales team test tomorrow?

Amplifa Product Amplifa helps industrial sales teams identify relevant target customers, triggers, and account priorities with AI — from CRM data, market signals, and ICP logic.

Why Pure Inbound Strategies Are Too Slow for AI Factories

I deliberately contradict a convenient thesis here: “If the customer is ready, they will come to us.” No. For AI-Factory projects, the customer often comes to the provider who picked them up at the right place six months earlier. The decision-making process does not begin with the inquiry. It begins with internal uncertainty.

A COO notices that scrap costs are rising. A plant manager is pressured to make OEE more transparent. An IT manager blocks cloud experiments. A managing director hears at a VDMA event in Frankfurt that competitors are working with Predictive Maintenance. There is no project yet. There is no budget yet. But there is tension. Anyone who then only waits for website conversions will see the market too late.

Anyone who still relies on a pure inbound strategy in 2026 will not have a pipeline worth the name in five years. Especially not in industrial SMEs. The winners will not be the loudest content machines, but the teams that read signals, form hypotheses, and enter conversations with technical relevance. Fewer newsletters. More timing.

This does not mean that outbound can again mean clumsy mass emails. Please no. I myself receive enough messages where my name is misspelled and someone wants to sell me “synergies.” Good outbound-oriented industrial sales looks different: understand the account, derive the technical situation, name a plausible trigger, formulate a short thesis, ask respectfully. A sentence like “We see in several Kuka and Beckhoff-heavy lines that local image evaluation is being re-evaluated precisely because of data residency” is not perfect. But it is a thousand times better than “I just wanted to introduce myself.”

AI Infrastructure: Costs, Context Windows, and the Hard Part

Now briefly technical, because otherwise only strategy remains. AI infrastructure does not only consist of GPUs. For industrial applications, latency, availability, network segmentation, data quality, model monitoring, interfaces, and operating costs matter. A model that runs cleanly in laboratory data can fail in production because the camera is positioned slightly differently, because oil mist changes the images, because a night shift uses different material, or because the PLC has a timing behavior that no one simulated in the test.

For generative AI in sales, another problem arises: context windows and costs. Large language models can process long documents today, yes. But anyone who indiscriminately dumps complete CRM histories, technical specifications, email threads, quote data, and service protocols into every prompt is not building intelligence, but a token heater. Token prices are falling, but bad architecture scales faster than good purchasing conditions. We repeatedly see setups where 80 percent of model costs are burned on irrelevant context.

The better architecture is usually unspectacular: normalize data, pre-filter signals, retrieve relevant excerpts, use the model only where language, classification, or pattern recognition truly provide added value. In sales, this means: Not every account update needs a large model. Sometimes a rule set is enough. Sometimes a small classifier. Sometimes you need an LLM because the customer's email reads between the lines: “We have a budget problem, but a greater risk if we do nothing.”

Precisely this sobriety is missing in many AI debates. Foxconn and Bull are not building relevant infrastructure because every SME will suddenly train its own Foundation Model. Honestly? Very few will do that. What is relevant is that industrial AI is moving closer to production and that Europe is gaining more control over integration, availability, and operation. This reduces friction. And friction often decides whether a sales team can even properly place an AI-related offer.

Amplifa for Industrial Sales For teams that want to translate technical buying signals like Edge AI, automation, service needs, and modernization projects into prioritized sales actions.

Personal Forecast: 2026 to 2029 Will Be Brutal for Sorting

My personal forecast for the next two to three years: The market will not sort itself by “AI yes or no.” It will sort itself by operational capability. Providers who sell AI as a demo will end up in pilot graveyards. Providers who package AI infrastructure, data flows, compliance, and service into an understandable commercial offering will win larger deals — even if they are not the cheapest.

I expect three concrete movements. First, tenders in mechanical and plant engineering will more frequently include questions about local inference, data residency, and model monitoring. Second, integrators will increasingly penetrate sales and push traditional manufacturers into consulting positions if they cannot explain their architecture. Third, existing customer data will become the most important growth lever, because AI-Factory projects rarely start from scratch. They start on existing lines, with old controls, with known pains.

Foxconn and Bull are not the only trigger here. But they are a visible signal. When a global manufacturing giant and a European infrastructure player jointly position AI hardware and automation platforms for Europe, it's not a marginal issue for IT departments. It's an indication of where industrial demand is heading. And demand is the material from which pipeline is built.

After the call with Andrea from Ulm, I printed out the Foxconn-Bull announcement again. Paper, not PDF. On the margin, in my handwriting, it says: “Not Hardware. Buying Criteria.” The coffee was long cold by then, that bright sawing sound came from production again, and Andrea texted me ten minutes later: “We need to rethink our target customer list, don't we?” Yes. That's exactly where it starts.