THE AI EDGE | Issue No. 3 | Tue, 7 July 2026 | Weekly AI intelligence for executives, on what's actually working in enterprise AI

Also published as The AI Edge on LinkedIn. Subscribe here →  The AI Edge

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Covering: Karp's AI Reckoning · EU AI Act Council Green Light · The Deployment Money Confirms It · Malta Financial Services Authority (MFSA) & Malta Gaming Authority AI Governance

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

## THIS WEEK AT A GLANCE

- Palantir's Alex Karp used a live CNBC interview on July 1 to call the AI industry "effing insane" - accusing leading AI vendors of charging enterprises for "tokens that create no value" and quietly "stealing weights and alpha." Palantir shares rose over 9% on the back of it. Strip out the theatrics and the underlying claim is a serious one: most enterprise AI spend is going into systems nobody can see inside or control.

- The Council of the EU gave final sign-off to the AI Act simplification package on June 29 — Official Journal publication is expected within weeks, and a new provision banning AI-generated non-consensual intimate imagery and CSAM was added. Article 50 transparency obligations land in 27 days.

- Microsoft, AWS, and OpenAI have now committed close to $8 billion, combined, to their own AI deployment arms — and the data explains why: only 5–6% of enterprises report substantial AI ROI, and 42% abandoned most of their AI initiatives last year, up from 17%.

- MFSA issued a Dear CEO letter on AI governance and the MGA confirmed it is drafting the first dedicated AI framework for iGaming operators — Malta's two principal regulators moved on AI oversight in the same month, ahead of most EU member states.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

# SECTION 1: THE BIG STORY

Karp's On-Air Meltdown Was Bad Television but a Fair Diagnosis

On July 1, Palantir CEO Alex Karp went on CNBC's Squawk Box and, in a nearly 20-minute stretch that outlets are now calling a "televised nervous breakdown," called the AI industry "effing insane." Strip away the delivery and the substance is worth an executive's attention. Karp said the CEOs he speaks with privately are "livid" because they are "paying for tokens that create no value," and accused leading AI vendors of effectively "stealing" customers' weights and alpha, the proprietary signal a company generates by running its own data through a model. He went further, questioning whether the West should "outsource the battlefield of this country to the consensus view in Silicon Valley," framing opaque, vendor-controlled AI as a sovereignty problem as much as a commercial one. Palantir's stock rose more than 9% that day.

Karp has an obvious commercial interest in this argument, Palantir sells the alternative. That does not make the diagnosis wrong. It matches, almost exactly, what this issue's own enterprise data shows: 42% of organisations abandoned most of their AI initiatives last year, and only 5–6% report substantial ROI. If the majority of enterprise AI spend genuinely is going into systems that produce fluent output without producing measurable value, "tokens that create no value" is not a soundbite. It is a fair label for the 2026 enterprise AI market.

The part of Karp's argument that deserves to survive the news cycle, independent of who said it: the enterprises actually getting value from AI are the ones that insisted on knowing what a system is doing with their data, retained control over the decision layer instead of handing it to a vendor's black box, and treated the move from prototype to governed production as the actual deliverable. That is a materially different requirement from "which model do we license." It is a question about transparency and control.

For boards, the practical takeaway is not to pick a side in a CEO's television appearance. It is to use the moment. Ask your AI vendors, this week, three questions Karp's rant effectively poses for free: can we see how this system reaches a decision, do we retain control of our own data and model outputs, and has this actually reached governed production or is it still a well-funded pilot. Vendors that answer plainly are only a few.

# SECTION 2: REGULATION & GOVERNANCE

The EU AI Act's Final Text Is Now Locked. The Clock on Article 50 Was Never Paused.

The Council of the EU gave its final green light to the AI Act simplification package on June 29, following the European Parliament's approval on June 16. Formal signature and Official Journal publication are expected within weeks, ahead of the August 2 applicability date.

The package confirms what was already understood: Annex III high-risk AI systems (recruitment, credit scoring, biometric identification, critical infrastructure) move to a December 2, 2027 deadline, and AI embedded in regulated products such as medical devices and lifts moves to August 2, 2028. The simplified SME compliance track now extends to companies with up to 750 employees and €150 million in annual revenue, widening its reach well beyond what the original Act covered.

None of this touches Article 50. Transparency obligations for AI systems that interact with customers or generate synthetic content become applicable in 27 days, on August 2. Any customer-facing chatbot, any AI-generated marketing image or voice, any synthetic content used in client communications must be disclosed as AI-generated. Fines run to €15 million or 3% of global turnover. Action item: if your organisation has not completed an inventory of every customer-facing AI touchpoint against Article 50's disclosure requirements, that inventory needs to be finished this month, not by August 2.

#SECTION 3: ENTERPRISE & INDUSTRY

The Deployment Money Is Confirming What Karp Said Out Loud

Three hyperscaler-adjacent organisations are now betting close to $8 billion, combined, on the exact problem Karp described. On July 2, Microsoft launched Frontier Company, backed by $2.5 billion and 6,000 industry and engineering specialists embedded at clients including the London Stock Exchange Group, Unilever, Land O'Lakes, and Novo Nordisk. AWS confirmed a comparable $1 billion commitment days later. Both follow OpenAI's DeployCo, launched in May with over $4 billion and McKinsey, Bain, and Capgemini as implementation partners. Notably, Microsoft's own framing of Frontier Company is model-agnostic, it will run OpenAI, Anthropic, Microsoft's own models, or open source, "without ceding control to any one of them." That is an admission, from a firm with a multi-billion-dollar OpenAI stake, that the deployment and control layer is where enterprises are actually stuck. But there are platform in the market satisfying this.

The underlying data explains the urgency. Only 5–6% of enterprises report capturing substantial ROI from AI, and 42% abandoned most of their AI initiatives in the past year, up sharply from 17%. 79% report meaningful adoption challenges. The pattern is consistent: pilots launch without predefined success criteria, so there is no basis for declaring success even when the technology performs as designed, and roughly half of proofs-of-concept never reach production because they were stitched together with manual workarounds that break under real load.

The 29% of organisations that do see significant ROI share three behaviours, none of them about the technology itself. They tie AI initiatives to revenue or cost outcomes before scaling. They name a single executive accountable for the financial result, a practice that triples the success rate versus committee ownership. And they measure outcomes before deploying rather than after, which alone correlates with a four-fold improvement in the odds of achieving ROI.

For boards evaluating any new AI deployment programme, hyperscaler-backed or otherwise, the dependency risk cuts both ways: these programmes embed the vendor's people inside your operations to accelerate delivery, which also means the vendor accumulates the institutional knowledge of how your AI runs. Contracts on data access, model portability, and knowledge transfer deserve the same scrutiny as the deployment timeline.

# SECTION 4: EMEA LENS

Malta's Two Regulators Moved on AI showing the way. That Is Not a Coincidence.

The MFSA issued a Dear CEO letter this month setting out supervisory expectations for AI adoption across Malta's financial services sector: boards and senior management are expected to take direct oversight of AI initiatives, with model validation, continuous monitoring, and data governance named as core requirements. In parallel, the Malta Gaming Authority confirmed it is drafting what would be the first dedicated AI governance framework built specifically for gaming operators, alongside its own 2026–2027 roadmap for applying AI to AML, player support, and financial compliance supervision. This is a strong signal that Malta intends to be a compliance leader on AI oversight rather than a follower, which matters for any operator using Malta as an EU base.

There is also a sovereignty thread worth pulling here. Karp's warning about outsourcing critical infrastructure to "the consensus view in Silicon Valley" is an American framing of an argument EMEA regulators have been making structurally for two years: control over the AI layer, not just access to it, is now a governance requirement, not a preference. MFSA and MGA's expectation of board-level oversight and data governance is the EU version of the same instinct.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

WATCH LIST

| 2 August 2026 | EU AI Act — Article 50 transparency obligations become applicable | 27 days |

| Ongoing, 2026 | MGA first dedicated iGaming AI governance framework — draft expected | Watch |

| 2026–2027 | MGA internal AI supervisory roadmap (AML, player support, compliance) | In progress |

| 2 December 2027 | EU AI Act — Annex III high-risk AI systems compliance deadline | 514 days |

| 2 August 2028 | EU AI Act — AI embedded in regulated products (Annex I) | 758 days |

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

## MY TAKE

Skip how Karp said it. Look at what he said: enterprises are paying for tokens that create no value because nobody demanded to see inside the system, or kept control of their own data.

That is the whole story of enterprise AI in 2026, in one line. The winners were never going to be the organisations with the best model. They are the ones that insisted on transparency into how a system reaches its output, and kept control of their own data and decision layer instead of renting someone else's black box.

I am not describing a hope. I have seen this built and running in production. The technology to demand a glass box instead of a black box exists today. Most boards are not short on will. They simply do not yet know this option exists, so they keep signing for the black box because nobody told them there was a different question to ask.

George

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

The AI Edge is published weekly by George Kakouras for informational purposes only and does not constitute legal, financial, or investment advice. Each edition covers enterprise AI deployment, strategy, and regulation for executives operating in EMEA