Conference Presentation, Panel
Clarifai, Oracle, TensorWave, Cherry & Vaire: Sovereign AI in a Borderless Tech World
- Infrastructure is the foundational prerequisite for Sovereign AI: Panelists agree that control over physical infrastructure, specifically power availability and physical space, must precede any software or value-based sovereignty goals.
- Peter Tomasik (TensorWave) identifies power and space as the two most critical constraints; without them, no stack can reside within a specific country or organization.
- Hakim Lumi (Oracle) argues that "sovereignty" extends beyond physical location to end-to-end control of the process, data, and decision-making logic.
- The scope of Sovereignty extends beyond territory to cultural and operational control:
- Don Barnetson (Credo) warns that hosting a foreign Large Language Model (LLM) locally does not confer sovereignty if the underlying data sources, training biases, and decision-making algorithms remain opaque or controlled by foreign entities.
- Rodolfo (VeriComputing) uses the analogy of hiring an employee whose decisions are influenced by a competitor to illustrate how using uncontrolled LLMs creates vulnerability to manipulation despite local data storage.
- Stranded assets are a primary financial risk in AI infrastructure investment:
- Hardware depreciation cycles (typically 5 years) are outpaced by performance doubling rates, creating a high risk of assets becoming obsolete as training clusters.
- The industry trend is moving from training clusters to inference clusters after approximately two years to repurpose Hopper-generation GPUs as Blackwell-generation clusters for training come online.
- Peter Tomasik notes that large-scale deployments require constant capital infusion ("constant money machine") for refurbishment once electricity costs rise or hardware becomes outclassed.
- The global AI race is effectively a binary contest between the US and China:
- Hakim Lumi states that Europe has already "lost the race" due to the immense capital scale required, citing a lack of major European companies in the top tier of AI infrastructure funding.
- Don Barnetson identifies the Middle East (specifically Saudi Arabia, UAE, Qatar) as a emerging third contender, leveraging energy wealth (electricity costs around $0.02/kWh) to build data center ecosystems.
- Hakim Lumi notes that the Middle East is positioning itself as a "data embassy" for other nations, utilizing its geographic centrality and energy capacity to become a global information refinery.
- Talent and energy are the primary differentiators for sovereign capability:
- Rodolfo emphasizes that talent is the single most important differentiator, with the "global talent war" making it difficult for regions like the EU to compete against US and Chinese capital.
- Don Barnetson argues that the Middle East's ability to build energy infrastructure is a more significant competitive advantage than their financial capital alone.
- Software and storytelling are critical gaps in current US and European AI strategies:
- Hakim Lumi observes that US companies focus heavily on technical progress, neglecting the cultural and narrative aspects necessary for global adoption and sovereignty.
- Don Barnetson notes a lack of a cohesive, government-backed national AI strategy in the US comparable to China's "AI by 2030" goal, leading to disjointed private-sector narratives.
- Peter Tomasik warns that European policymakers often fail to commit to the long-term operational costs of data centers, leading to "stranded assets" after initial capital investment.
- Pragmatism currently overrides sovereignty in enterprise adoption:
- Hakim Lumi reports that while customers publicly prioritize sovereignty and security, market reality drives them toward cheaper, "black box" open-source or global solutions if they perform better.
- Matt Zehler (Clarify) clarifies that sovereignty requirements are use-case dependent; while general coding or chat applications are location-agnostic, military and national security applications demand full control over the stack.
- Security risks are amplified by the complexity of LLMs:
- Hakim Lumi warns that LLMs act as new attack vectors capable of exfiltrating sensitive data through natural language interfaces, necessitating "security by design" and middleware to control access based on user roles.
- Future Outlook and Strategic Advice:
- Hardware: Peter Tomasik advises against single-supplier dependency and recommends diversity in silicon (e.g., AMD, NVIDIA) to ensure long-term viability.
- Network Architecture: Rodolfo and Don Barnetson emphasize that network "plumbing" is the second most expensive data center cost and requires careful architectural choices to match US hyperscaler efficiency.
- Policy: Don Barnetson suggests lobbying for policy changes, including talent importation, union restrictions, and capital gain tax reductions to improve competitiveness.
- Investment: Nadia (Cherry Ventures) highlights a strong interest in pre-seed to Series A investments in energy, photonic chips, and AI orchestration software.