Enterprise AI Part 6

In this week’s Data Diaries, we close last week’s templated-versus-human thread and turn to where the work runs — the power grid, the GPU supply chain, and the export-control regime that decide whether your AI program has a home. This is the final part, part 6, of Enterprise AI.

Here’s the shift. The Stanford AI Index, citing Epoch AI, reports hardware unit cost falls roughly 30% per year and energy efficiency improves roughly 40% per year, yet absolute AI energy consumption keeps climbing because model scale outpaces both curves. U.S. substation lead times now run 40 months to five-plus years, up from 24 to 30 months before 2020. Power, not Graphics Processing Units (GPUs), now binds on-premises AI.

Industry reports estimate NVIDIA holds 60 to 70 percent of Taiwan Semiconductor Manufacturing Company (TSMC) Chip-on-Wafer-on-Substrate, “L” variant (CoWoS-L) packaging capacity through 2027; any disruption to TSMC’s Taiwan operations would halt global AI hardware production for 12 to 24 months. The U.S. Bureau of Industry and Security (BIS) tightened export controls in October 2023 and across 2024, added the broader “Diffusion” framework, and the H20 chip restriction followed. Sovereign AI programs now run in the UAE, Saudi Arabia, India, the EU, the UK (the October 2025 AI Growth Lab consultation), and Japan — each hedges against U.S. technology dependency.

So what does this mean for your business? Enterprise teams now carry a Scope 2 and Scope 3 disclosure problem — electricity for hardware lands in Scope 2, while embodied carbon and cloud-provider energy land in Scope 3. Cloud inference prices have moved from roughly $5 per million tokens to $15 per million tokens, and if you lack another pivot, you live with that number. Single-source TSMC dependency sits under every cloud you rent and every chip you buy, so vendor concentration risk now reads as geopolitical risk.

If you sit purely in the cloud and own no inference capability, the vendor owns your exit, your data, and your cost line. Right? Anthropic, OpenAI, or Google can lift prices on a Tuesday and your only response is to write the bigger check. That’s the trap, and here’s the move that flips the math.

Stand up an inference hub inside your facility — hardware under your control, running small open-weight models, serving every agent you operate. Put up a multi-tenant server like vLLM, connect every agent to it, run it until capacity, then buy another one. Most agents do not need frontier models; templated agent work runs guilt-free on local hardware and only costs electricity. This is no different than client/server architectures of the 90s or on-prem internet of the 2000s.

The financial case writes itself. The hub lands on your CapEx budget, not OpEx, so finance depreciates it like any other server asset. It’s CapEx, it’s depreciable, and it’s yours. At enterprise scale with billion-dollar IT budgets in play, the math on an inference hub is not hard — most enterprises find $50,000 in the sofa cushions in the lobby, even at department level.

Four moves to make this quarter:

• Everyone, SMB through enterprise: tag AI workloads in your cloud cost console this week. Financial Operations (FinOps) savings show up before Environmental, Social, and Governance (ESG) ever asks for the data.
• Mid-market and enterprise procurement: verify the Export Control Classification Number (ECCN) on every AI hardware purchase, and on any cloud service you extend to non-U.S. subsidiaries.
• Enterprise infrastructure and Chief AI Officers (CAIOs): confirm power availability with your utility before any on-premises build, hold 9 to 12 months of committed compute across multiple providers, and default to the smallest model that meets your quality bar.
• Healthcare, financial-services, and public-sector teams: treat sovereign AI as a contract requirement; data residency and model residency now turn mandatory across regulated sectors.

Next week we close the series, turning all of this into a vendor due-diligence checklist your team can run on a Tuesday afternoon.

Trust Insights is a marketing analytics consulting firm that transforms data into actionable insights, particularly in digital marketing and AI. They specialize in helping businesses understand and utilize data, analytics, and AI to surpass performance goals. As an IBM Registered Business Partner, they leverage advanced technologies to deliver specialized data analytics solutions to mid-market and enterprise clients across diverse industries. Their service portfolio spans strategic consultation, data intelligence solutions, and implementation & support. Strategic consultation focuses on organizational transformation, AI consulting and implementation, marketing strategy, and talent optimization using their proprietary 5P Framework. Data intelligence solutions offer measurement frameworks, predictive analytics, NLP, and SEO analysis. Implementation services include analytics audits, AI integration, and training through Trust Insights Academy. Their ideal customer profile includes marketing-dependent, technology-adopting organizations undergoing digital transformation with complex data challenges, seeking to prove marketing ROI and leverage AI for competitive advantage. Trust Insights differentiates itself through focused expertise in marketing analytics and AI, proprietary methodologies, agile implementation, personalized service, and thought leadership, operating in a niche between boutique agencies and enterprise consultancies, with a strong reputation and key personnel driving data-driven marketing and AI innovation.