Cerebras $2.5B: physical AI changes the rules for SMEs

The timeline of an exit the market didn't see coming

In May 2026, TechCrunch reported the details of Cerebras Systems' $2.5 billion exit. The Eclipse fund, founded by Lior Susan, is among the main beneficiaries of the deal. But the story starts much earlier.

About ten years ago, Susan began investing in companies building technology for the physical world. It was a lonely position. Most venture capital was looking at software, SaaS and digital marketplaces. Eclipse, by contrast, bet on chips, sensors, robotics and tangible computing infrastructure.

Cerebras has become the most visible proof of that thesis. The company developed chips of extraordinary size, the Wafer Scale Engine, capable of running large AI models with low latency. It also built complete systems, not just raw silicon.

Cerebras architecture: why silicon matters more than software

To understand the value of the deal, it helps to look at the technical architecture. Cerebras chips don't follow the traditional GPU cluster paradigm. Instead, they integrate thousands of cores on a single silicon wafer.

As a result, communication between compute units happens without the bottlenecks typical of inter-GPU networks. This translates into concrete advantages on specific workloads: language model training, high-frequency inference, complex physical simulations.

According to analysis published by Gartner, by 2027 more than 60% of companies adopting AI in production will have to make structured decisions about inference hardware. So choosing a chip is no longer something to hand off to the IT department: it becomes a strategic decision.

For SMEs building in-house AI pipelines in particular, such as predictive analytics, document automation and product recommendation, the availability of specialized infrastructure changes the make-or-buy math.

Winners and losers in the AI hardware ecosystem

The Cerebras exit reshuffles some hierarchies. Among the clearest winners are Eclipse and its LPs, but also the entire segment of physical AI investors. The deal validates an investment thesis that for years stayed marginal compared with software-first.

Among those who need to rethink their positions are general-purpose cloud computing providers. Companies like AWS, Google Cloud and Azure built their advantage on horizontal flexibility. But the vertical specialization of AI hardware creates niches where traditional cloud isn't competitive on either cost or performance.

Likewise, traditional GPU makers, NVIDIA first among them, need to watch the rise of alternative architectures. Even so, NVIDIA keeps an enormous ecosystem advantage thanks to CUDA and its installed base. The duopoly isn't in question in the short term.

To dig deeper into the competitive dynamics in AI hardware, the McKinsey State of AI report offers a structured look at the tensions at play among the industry's giants.

The SHM Studio take: what changes for companies that aren't hyperscalers

At SHM Studio, we work every day with Italian SMEs that are integrating AI into their processes. So we look at this story from an operational angle, not a speculative one.

Here's the key point: physical AI infrastructure is becoming accessible beyond the big data centers. This happens through two main channels.

• Specialized cloud: providers like CoreWeave or Lambda Labs offer hourly access to Cerebras chips and NVIDIA alternatives. An SME can test workloads on different architectures without buying hardware.
• Hybrid on-premise models: some manufacturing and logistics companies are evaluating local deployments for latency and data sovereignty reasons. In this scenario, the choice of chip becomes part of the overall architecture project.

So the question for an SME isn't "Should I buy Cerebras chips?" The question is: "Does my AI roadmap for the next 24 months call for a rethink of my computing infrastructure?" In most cases, the answer is yes.

For companies building a structured digital presence, our AI services include a technology assessment phase that also considers infrastructure choices. In the same way, our approach to digital marketing always builds in intelligent automation.

Operational implications for Italian SMEs

Turning financial news into concrete action is something SMEs often struggle to do in-house. So let's outline a few practical directions.

First, map your current and planned AI workloads. Not every use case needs specialized hardware. For example, a text classification model at medium volumes runs perfectly well on standard cloud instances. Real-time inference on industrial sensor data, on the other hand, can benefit from dedicated architectures.

Next, keep an eye on how prices evolve in the specialized cloud market. Competition among providers is pushing costs down. Solutions that seemed out of reach 18 months ago now fit the budget of companies with 50-200 employees.

Finally, it's worth setting up governance for AI technology choices. That means not leaving infrastructure decisions to the IT department alone, but involving senior leadership and, where there is one, the CFO. The implications for TCO (Total Cost of Ownership) are significant.

For those running digital campaigns with automation components, choices around Google Ads and LinkedIn Ads are also moving toward heavier use of predictive models. So the quality of the underlying infrastructure directly affects campaign performance.

The work still ahead: where Eclipse goes after Cerebras

According to TechCrunch, Lior Susan sees the Cerebras exit as only the beginning of the physical AI thesis. Eclipse has other companies in its portfolio working at the intersection of hardware, robotics and artificial intelligence.

This signal matters for understanding where capital will concentrate in the coming years. The "physical world AI" thesis also overlaps with trends already under way: the digitalization of manufacturing, logistics automation, computer vision systems for retail.

For Italian SMEs, many of which operate in exactly these sectors, the message is clear. AI is no longer just a topic for software houses or digital startups. It matters to anyone who produces, distributes and sells in the physical world.

Our SEO and copywriting work for B2B SMEs takes this shift into account. So does our approach to web development, which increasingly integrates data-driven personalization.

Next moves: three questions to guide you

Let's close with a simple framework: three questions every SME should ask after reading this news.

• Which business processes could benefit from real-time AI? If the answer includes physical operations (production, warehousing, field service), then computing infrastructure becomes relevant.
• Does our current cloud provider offer access to specialized AI hardware? If not, it's worth exploring alternatives before your AI roadmap runs into technical limits.
• Do we have governance for our AI technology decisions? If choices are made case by case without a big-picture view, the risk of architectural inconsistency grows over time.

To explore these topics with a consulting approach, you can contact the SHM Studio team. Our blog also regularly publishes analysis on AI, digital infrastructure and strategies for Italian SMEs.