I spent much of Q2 in rooms with the people building, funding, and buying AI at scale: Oslo for the INSEAD Alumni Forum, Munich for the Entrepreneurship Forum, Paris for the INSEAD AI Forum, where Turing winner Yann LeCun and Nobel laureate Philippe Aghion both spoke. I’ve recapped my learnings here -https://www.linkedin.com/posts/poojajain19d_powerup-ai-q2-learning-recap-ugcPost-7479852796029153281-W-ug/?
Every room circled the same question: who controls the infrastructure you are building on, and what happens when that control is exercised without warning?
Q2 delivered three concrete reasons to ask.
#1 Agents changed the unit of work
For two years, using AI meant prompting. You asked, it answered, the work stayed with you.
Q2 broke that assumption. The unit of work became the workflow: research the account, pull the CRM history, draft the follow-up, flag what needs judgment, stop. The system plans, uses tools, checks its own output, and returns when done or stuck.
The tooling landed in a single quarter: Microsoft Agent Framework 1.0 (April 3), Google's Gemini Enterprise Agent Platform wired into SAP, Salesforce, ServiceNow, and Workday (April 22), Anthropic's Cowork, and comparable moves from OpenAI and DeepSeek.
The consequence for how you run things: AI just got promoted from tool to infrastructure. Infrastructure has to be governed, measured, and protected against failure. The rest of this issue is what happens when it changes underneath you.
#2 The spend is real. The measurement is missing.
Uber spent its entire 2026 AI budget in four months. Q1 R&D hit $951 million, up 17% year on year. COO Andrew Macdonald in May: "That link is not there yet," meaning he cannot connect the spend to measurable product improvement.
Every large company sits in the same position. Uber just said it out loud. The hyperscalers can afford unclear returns. Combined CapEx across Amazon, Microsoft, Alphabet, and Meta is tracking toward roughly $725 billion in 2026, up 77% over 2025.. They are buying infrastructure dominance and can wait years. A mid-size company cannot.
Forrester found 25% of planned enterprise AI spend already slipped to 2027 under financial scrutiny. Grant Thornton surveyed 950 business leaders: 78% would fail an independent AI governance audit within 90 days. Spending without measurement gets very uncomfortable the moment a board asks what the money bought.
#3 19 days without the model you built on
June 9: Anthropic launched Claude Fable 5. June 12: a US export-control directive, and Anthropic suspended global access because it had no way to filter by nationality in real time. Stripe, Hebbia, and Mozilla watched live workflows pause. The model returned July 1, 19 days later, with a classifier that quietly reroutes some legitimate requests to an older model.
If a meaningful share of your team's output runs through a model you do not control, an unplanned 19-day suspension is a business continuity event.
Satya Nadella's June test is the one to keep: can your company swap out a generalist model without losing the expertise built into your own systems? Most organizations cannot answer that today. Read here: https://x.com/satyanadella/status/2066182223213293753?s=20
Big Question: Who controls AI?
Yann LeCun was speaking INSEAD AI Forum and he clearly named the industry's contradiction: the largest AI companies preach democratization while restricting their most capable models because ordinary people supposedly cannot be trusted with them. His framing: that is putting yourself on a pedestal. He compared it to the Ottoman Empire's long ban on the printing press, done partly to control the dogma and partly to protect the scribes whose livelihood the press threatened.
Bottomline: Single-provider dependency belongs on your risk register.
China is live proof that constraint forces independence. Zhipu's GLM-5.2 (June 13) beats GPT-5.5 on SWE-Bench Pro (62.1% vs 58.6%), runs on Huawei silicon, carries an MIT license, and costs roughly a quarter of Claude Opus. DeepSeek V4 needs 27% of its predecessor's compute. For Western enterprises the uncomfortable point stands: critical capability built on infrastructure you do not control, at investor-subsidized prices, deserves a clear-eyed look.
The knowledge layer is the only sovereignty you can build
A year ago, client conversations were about model choice. That question is settled and commoditizing; frontier models cluster together and a capable open one is never far behind.
What decides whether AI output can carry real work is the knowledge layer: documented workflows, domain expertise, evals that test a model for your context, institutional judgment made queryable. With that layer in place, swapping a model is an engineering decision. Without it, every model change starts you over. That is what Nadella's test measures.
Based on everything I've learned over the past few months, I've launched AI Ops Blueprint 2.0. It's a two-week engagement that maps where AI fits into your operations, identifies key dependencies, and highlights where your knowledge layer needs to be built.
Reply to this email if you'd like to learn more.
Until next time,
Pooja
PS: If you know a CFO or COO wrestling with "we are spending real money on AI and cannot show the board what it bought," forward this. That is the conversation I am in most weeks.