Compute is not your bottleneck: what the AI infrastructure race means
In short
The United States controls roughly three quarters of global AI compute. For your SME, that is not a threat: compute is rented per query for cents and gets cheaper every quarter. Your real bottleneck is not infrastructure, but selecting and orchestrating the right processes.
The US is building more compute, and your SME benefits
The global race for AI compute looks like a problem at first glance. The United States increased its share of shipped AI chips from around 60 to 63 percent in 2022 to 2024, to 72 percent in 2025, and to 78 percent in the first half of 2026. According to Samaritan Research, 72 percent of the 27.4 million H100 equivalents delivered between 2022 and mid-2026 went to the US. For your SME, however, this is not a threat but a structural advantage: compute becomes cheap rented infrastructure you do not need to own.
The numbers behind the race
72%
US share of shipped AI chips from 2022 to mid-2026, according to Samaritan Research
75.3%
US share of operational AI compute in October 2026, according to CUDO AI Observatory
The AI race is becoming a battle for compute.
Switzerland holds 0.9 percent of operational AI compute according to CUDO. That is not a weakness for a small high-cost country; it is reality. What matters is that this gap does not affect your daily operations as long as you can rent compute per query.
Compute is rented infrastructure, not a strategic asset
Not long ago, AI infrastructure meant owning servers, GPU clusters, and a team to run them. Today you rent compute per query, often for cents, and prices continue to fall. What used to be a capital project is now a variable operating expense. There is no rational reason for an SME to buy high-end GPUs just because the US is expanding capacity. You benefit from billions in data center investment without investing yourself.
Your real bottleneck is application and orchestration
The real work is not access to compute. It is the last mile: which processes do you automate first? How do you bring data from SAP, Abacus, or Microsoft 365 into a clean AI workflow? Who ensures compliance with GDPR or Swiss revDSG? This orchestration determines whether an AI project creates value or becomes an expensive experiment. You do not need to build an internal AI team; a forward-deployed engineering partner can implement processes directly in your business.
The last mile in Swiss reality
Swiss companies face practical hurdles: business applications like Abacus or Bexio are deeply embedded, data often sits in M365 or local systems, and revDSG demands clear accountability. Whoever introduces AI must connect these systems, clarify data quality, and take operational responsibility. That is doable, but it is the actual work. Focusing only on compute misses exactly this part. For a strategic view on how Swiss companies should position themselves, listen to the Swiss AI Podcast.
In short
Compute is a commodity for your SME. Your differentiation lies in process selection, data integration, and operations. Exactly this layer can be bought as AI outsourcing.
The race works for you: how to use it
- Choose two or three processes with high manual effort and clear data access.
- Check integration with Abacus, Bexio, or Microsoft 365 and clarify data quality and revDSG obligations.
- Use external AI specialists for orchestration and operations instead of building an internal team.
- Start with a limited pilot project that shows measurable value within a few weeks.
Compare: owning compute vs. renting it
- Own GPU servers
- High fixed costs, rapid obsolescence, in-house operations, rarely economical for SMEs.
- Cloud compute (pay per use)
- Variable cost per query, instantly scalable, no hardware risk, requires technical setup.
- AI outsourcing as an external division
- Bundles compute, orchestration, integration, and operations; focuses on process impact rather than infrastructure.
- Self-orchestration with an internal team
- High recruitment and knowledge burden, long lead time, not viable for many SMEs.
Practical start
Begin with one process you already know well, such as invoice intake or proposal generation. Infrastructure is not your risk; process selection is.
Frequently asked questions
- Does my SME need its own AI compute?
- No. For typical SME use cases, compute is a variable rented resource billed per query. Owning GPU servers only makes sense in special cases with very high, sustained volume or strict sovereignty requirements.
- What does the US lead in AI compute mean for Swiss SMEs?
- The lead makes AI globally cheaper and more capable. For your SME, what matters is not who owns the servers, but who builds the best applications and orchestration.
- What is meant by orchestration?
- Orchestration includes process selection, data integration with systems like Abacus, Bexio, or Microsoft 365, quality assurance, and ongoing operations. This is the work that determines value.
- How do I start without an internal AI team?
- You can use AI outsourcing as an external division: a specialized team handles setup, integration, and operations while you retain business control. Start with a clearly defined pilot process.
- What role does revDSG play in AI projects?
- revDSG requires clear accountability and data protection for personal data. With cloud services, you need binding contracts, suitable hosting locations, and documented processes. This is solvable but must be clarified early.
Sources
- Alvin Foo auf LinkedIn: «The AI race is becoming a battle for compute» (Inspiration)
- Samaritan Research: AI Chip Imports Explorer (2026)
- CUDO AI Observatory / Epoch AI: Share of operating AI compute by country (Okt 2026)
- Carnegie Endowment: How to Build the Future of AI in the Free World (2025)
- Roland Berger: Data Center Study 2026 — The global build-out race
Chris Jon Graf · AI Strategist & CEO
For over 15 years, Chris Jon Graf has built productive Swiss software platforms and advised executive teams on the strategic deployment of artificial intelligence. Would you like to evaluate this topic for your company?