The US and China are building two incompatible AI stacks, and the companies still mixing both are on a timer. Plus the EU AI Act slips, PE keeps underestimating AI, and the herd mentality eating VC.

This week from the director's seat: the US and China are splitting the AI world into two incompatible stacks, the EU AI Act slips (but not where it counts), private equity keeps underestimating AI, and the herd mentality eating venture capital.

The Great Divide: how the US and China are splitting the AI world

BCG's new analysis describes something more structural than competition: the US and China are building two increasingly incompatible technology stacks, and companies that assumed they could keep mixing both are on a timer.

The map is uneven by design. The US holds its lead where frontier work gets done: top-tier talent, capital and IP. China is running a different race, pushing cost-optimized models into adoption at national scale, with the data and compute build-out to match. The middle powers are hedging in different directions: the EU is buying sovereign compute, Japan is aligning with Washington, and India keeps a foot in several ecosystems at once.

Across talent, IP, data, compute, energy and capital, only two countries cover most of the map. Everyone else is choosing which orbit to join.

Your AI stack used to be a procurement decision. It is becoming a statement about where your company can operate.

Boardroom takeaway. BCG's advice is to build the geopolitical muscle now, while mixing is still possible, because the stack you sit on will increasingly determine which markets you can serve and how exposed you are to political shocks. Ask management for a dependency map: which models, chips, clouds and data flows sit on which side, and what a forced separation would cost. If nobody in the room owns that answer, that is the finding.

Don't give away your edge

Alex Karp accused the big AI labs of "stealing the weights and alpha" of the companies that use them, and Palantir published a nine-point manifesto on AI sovereignty. The question every board should be asking: when your company uses these tools, who actually gets smarter, you or the vendor? Read the full piece, with a director's checklist.

Lobbying around the EU AI Act pays off, almost

The EU's Digital Omnibus package was formally passed in early July, and with it the AI Act's high-risk regime officially slips. Stand-alone high-risk systems under Annex III, which covers recruitment, credit scoring, education and similar use cases, now have until December 2, 2027. High-risk AI embedded in regulated products under Annex I, such as medical devices and machinery, moves to August 2, 2028. After more than 110 EU businesses lobbied for a pause, the pressure visibly worked.

Look at what did not move. The transparency obligations still land on August 2, 2026. And the outright bans held and got stricter, adding new prohibitions on AI that generates non-consensual intimate imagery or child sexual abuse material. Brussels gave ground on paperwork, not on substance.

Lobbying bought the industry sixteen months. It did not buy your board a defense.

Boardroom takeaway. Treat the new dates as sequencing, not relief. The work that makes December 2027 achievable, finding every AI system in the organization and classifying which ones are high-risk, does not get easier with time, and the inventory is also what the August transparency duties depend on. Board question: which of our systems would count as high-risk under Annex III, and who owns that inventory today?

Private equity underestimates AI

Varick Agents pulled apart why AI value creation stalls inside PE portfolios, and the diagnosis is not a lack of effort. It is fragmentation. In their example, a portfolio of just four companies needed 110 separate AI deployments, because every portco runs its own variants of accounts payable, reporting, close, collections and procurement, split by region, team and entity.

The fix they describe is a two-step consolidation: standardize workflows inside each company first, then group companies that share the same logic. In their accounts payable example, more than 100 separate transformations collapsed into 3 reusable agents.

The part fund directors should sit with is the return math. Standardization does not just expand EBITDA. It supports higher exit multiples with growth-oriented buyers who pay for clean, repeatable operations, and it shortens holding periods, because a transformation pattern proven in one portco redeploys across the rest in a fraction of the time.

The portfolio does not need more AI pilots. It needs fewer, reused.

For boards and investors. Ask whether the value creation plan prices AI at the portco level or the portfolio level. One is a cost line. The other is where the multiple lives.

The herd mentality eating VC

Dan Gray, research lead at Odin, argues that venture capital's reliance on consensus narratives is eroding the one thing the asset class is priced on: intellectual diversity. Odin has published a research directory of 135 papers spanning four decades, from the 1980s through 2026.

Two mechanisms do the damage. Processing fluency rewards simple stories, so the cleanest narrative beats the truest one. And the spiral of silence keeps non-consensus views unspoken wherever disagreement is expected, which in venture is everywhere. The returns tell the rest: the research points to outsiders outperforming in aggregate, while the ZIRP-era convergence of fund strategies produced a generation of disappointing vintages.

An asset class that gets paid for non-consensus bets cannot run on consensus narratives.

For fund directors and LPs. At the next fund board or LPAC, ask where the fund's thesis genuinely departs from the market. If the answer sounds like every other deck this quarter, you are underwriting beta dressed as alpha.

Also worth a minute

KPMG hands the Chief AI Officer question to HR. Two years ago, 11% of large enterprises had a Chief AI Officer. Today it is 26%, and KPMG argues the strongest claim on the role now belongs to the CHRO, not the CIO, because agents are becoming a category of worker. One enterprise in their analysis runs a 3:1 ratio of agents to employees. If agents are on your org chart, ask who manages them.

OpenAI wants Washington on the cap table. Per the Financial Times, OpenAI has proposed handing the US government a 5% stake, worth roughly $42.6 billion at its last valuation, with Sam Altman suggesting the other leading labs contribute similar stakes into a sovereign-wealth style vehicle. The talks are described as preliminary. For directors, the signal matters more than the deal: your most critical AI vendors are becoming politically entangled counterparties, and vendor concentration now carries a geopolitical line item.

Frequently asked questions

How are the US and China splitting the AI world?

They are building two increasingly incompatible technology stacks. The US leads on frontier talent, capital and IP; China pushes cost-optimized models into adoption at national scale. Middle powers are hedging: the EU buys sovereign compute, Japan aligns with Washington, and India keeps a foot in several ecosystems at once.

What does the US-China AI divide mean for companies and boards?

The AI stack you sit on, your models, chips, clouds and data flows, increasingly determines which markets you can serve and how exposed you are to political shocks. Boards should ask management for a dependency map, and what a forced separation between the two stacks would cost.

This is a roundup from The AI Leadership Edge, our newsletter for investors and boards. Powered by BoardLens: your board materials, read against thousands of outside signals through your lens, never trained on and auditable to source. Independent directors use it on one board; GPs across the whole portfolio. Start free.

About BoardLens

Private market intelligence for boards and portfolios. Your AI analyst for board and post-investment work.

BoardLens gives directors and investors a single secure workspace across every board they serve, while keeping each company's information completely separate and confidential. It reads your board materials into a clear, cited brief, tracks commitments and follow-ups, monitors the portfolio for drift, and keeps a governance calendar of meetings, covenants and reporting deadlines.

Governance-grade by design: your data is never used to train AI models, enterprise agreements and DPAs are in place with every underlying provider, and all data is encrypted in transit and at rest, in an isolated workspace per company.

Start free, with a demo company already set up so you can explore it without uploading your own materials. www.boardlens.ai

The Great Divide: How the US and China Are Splitting the AI World