Data Pains → Business Gains

May 28, 2026

So, What Actually Happened?

So, Thursday morning, and the whole week the AI conversation finally moved past which model to buy and started asking where the thing actually plugs in. We scanned 190,000 articles this week so you don't have to. Snowflake committed $6 billion to AWS and bought Natoma to wire the agentic enterprise, DigitalBridge and ArcLight merged into one asset manager sitting on top of the power-AI stack, and a hard look at the rollout reality found only 12% of AI pilots reach production. Meanwhile a Federal Reserve Governor named AI a macro variable in the same kind of speech the Fed once reserved for tariffs and energy.

The Bottom Line: The model is the record. The integration, the power, the conversion to production, that is the venue. Everyone bought the same record this week. The money rushed into the venue.

The Tracks That Matter

1. Snowflake Buys Agentic Plumbing With A $6B AWS Commit

Here's the deal that named the week. Snowflake announced intent to acquire Natoma, a secure-connectivity layer for AI agents, while committing $6 billion to AWS over five years to underwrite the agentic enterprise stack. The same day, Jefferies put Snowflake on the ”huge winner” list on the back of the same agentic narrative. Three signals, one bet: identity and connectivity for agents are the next moat.

The shape of the bet rewards reading. Natoma sells the boring middle layer that lets AI agents talk to your systems without burning a hole through your security policy, single sign-on for software that acts on your behalf. Snowflake's $6B to AWS is the long capacity reservation that lets that middle layer actually run at enterprise scale, not just demo well. And the Jefferies read names what the buyers actually care about: not the model, but the secure path between the model and the data it needs to act on. The package only makes sense if you accept that agentic AI's bottleneck is no longer intelligence, it is integration.

This is where the agentic story is finally landing. For a year the pitch was ”agents that can do things.” The honest version, the one procurement signs off on, is ”agents that can do things without leaking, without breaking SSO, without taking the audit committee hostage.” Snowflake's bet, identity-and-pipes plus reserved compute, is the unglamorous version of agentic AI that actually ships through enterprise IT. The model is the headline. The plumbing is the purchase order.

Here's what works: Before you approve any agentic AI proposal, ask which system handles agent identity and connectivity, and where the capacity to run it physically lives. If the answer is ”we'll figure that out later,” the project will ship a demo and stall before production. Buy the plumbing first, then plug in the model.

2. DigitalBridge And ArcLight Merged Where Power Meets AI

This is the M&A that names which stack runs the next decade. DigitalBridge and ArcLight announced a strategic combination to form a leading alternative asset manager sitting precisely at the convergence of power, AI, and digital infrastructure. Not a model lab. Not a chatbot. The asset class is the wiring under all of it: the megawatts, the racks, the fiber, the cooling.

The thesis is plain in the release: the firms that own the physical foundation will set the terms of the AI buildout. In the same corpus this week, Grundfos's CEO said water and energy will decide whether Europe can scale AI at all, and the AI data centre boom raised the stakes on design and environmental standards for the firms certifying these builds. Three stories, one read: the bottleneck moved from algorithms to amperes, and the firms positioning to own that constraint are pricing themselves as a category.

For your AI strategy the consequence is concrete. The compute, the power, the cooling, the certification, all of it is now the limit, not the model. The infrastructure operators are positioning to be the gatekeepers, and the prices and queues they set will shape what you can ship two years from now more than which lab you partnered with last quarter. The board that asks about model partnerships and skips capacity and power is reading a 2024 map.

Here's what works: Add a line to your AI vendor reviews about where the compute physically lives, who supplies its power, and how long the queue is. If your provider is buying capacity on the spot market, your roadmap is on the spot market with them. Lock in a multi-region, multi-supplier capacity story before the buildout finishes.

Try It Yourself

3. Only 12% Of AI Pilots Make It To Production

Here's the receipt for two years of AI breathlessness. A fresh look at enterprise rollouts found the cross-industry conversion rate from pilot to production is just 12 percent, with 46 percent of proofs-of-concept abandoned before production and 42 percent of companies scrapping most AI initiatives in 2025. Eight in ten pilots die quietly somewhere between the demo and the deployment.

What is killing them is not the model. The same analysis names the cause: pilot purgatory is now overwhelmingly an organizational and cultural problem, not a technical one. Data is not ready, success is not defined, the production system has nowhere to put the output. And the clock is loud: the EU AI Act high-risk window opens August 2, 2026, with compliance costs of $8 to $15 million initial and $1 to $5 million annual on large organisations. Pilots that have not crossed by then face a regulatory bar on top of the budget bar.

So we end up at a calibration moment. The optimistic AI valuations of the past two years quietly assumed deployment would follow demo. The data now shows it does not, four pilots in five never make it. Anyone planning their AI roadmap on assumed conversion rates is planning a chart that does not survive contact with their own ops review.

Here's what works: Pull your AI portfolio into a single list and mark each line ”pilot,” ”production,” or ”dead.” If your production column has fewer than two in ten, you have a conversion problem, not a model problem. Fix the handoff, the named owner, and the success metric before you green-light the next pilot.

4. Hippocratic AI Hit 10 Million Patient Calls At 99.9% Safety

In a week mostly about plumbing, here is a deployment story with actual results. Hippocratic AI scaled to 10 million patient calls at 99.9% clinical safety on DigitalOcean's AI-native cloud powered by NVIDIA Blackwell Ultra GPUs. Not a demo, not a pilot. Ten million phone calls to actual patients, with a clinical safety rate three nines deep, all measured.

What makes this different is the layer stack it actually proves out. A clinical-safety floor at 99.9% on 10M calls is the kind of operational evidence that survives a real diligence review, exactly what the AI valuation conversation is now pushing toward, milestones over multiples, measurable outcomes over claims. And it lands on a specific compute substrate, DigitalOcean cloud plus Blackwell Ultra, which tells you something about where high-throughput clinical AI is actually running now. Not the obvious hyperscaler default.

Read this next to the pilot-purgatory number from story three and the picture sharpens. Most healthcare AI never gets past the demo. The ones that do are focused vertical companies running on alternative cloud infrastructure with safety as a hard SLA, not a slide. The bottleneck on healthcare AI is not innovation, it is the discipline to actually measure clinical outcomes on real calls before claiming the win.

Here's what works: In any vertical AI proposal, demand the same shape of evidence Hippocratic just shipped: volume number, error rate, infrastructure named, time window stated. ”AI for healthcare” is not a number. ”10 million calls at 99.9%” is one. Ask for the second shape every time.

5. The Fed Just Said AI Is The Macro Story Now

Federal Reserve Governor Lisa Cook stepped to the mic this week and named AI as the macro story shaping productivity, labor, and the economy outright. The full speech is unusually direct from a Fed Governor: AI is no longer a tech-sector phenomenon, it is a central-bank-level variable shaping productivity growth, labor market dynamics, and inflation forecasting.

It lines up with another move you should not miss. The OpenAI Foundation committed $250 million to research the economic changes from AI, a quarter of a billion dollars to study the labor consequences of the technology the same parent company is shipping. When the lab funds the labor-impact research and the Fed names AI as a macro variable in the same week, the signal is that the political and economic debate is shifting from ”will it” to ”how fast and to whom.”

For enterprise strategy this is the part that actually matters. Your headcount plans, your productivity assumptions, your training budget, they are now operating inside a macro narrative the central bank is actively monitoring. The ”we'll figure out the people part later” stance has a shrinking window before the macro frame catches up to the operating frame.

Here's what works: If your AI strategy does not have a named human-impact line item, re-skilling, role redesign, workforce communications, add it this quarter. The macro conversation just landed at the Fed level, and the boards that have not asked yet will start asking by Q3.

6. The Cyber Extortion Economy Just Got Its Field Manual

Hidden in the security wires this week, a piece worth printing for anyone with AI in production. Palo Alto's Unit 42 published a deep look at the evolving cyber extortion economy, and the picture is industrialised: not lone-wolf ransomware anymore but a market with specialisation, supply chains, and pricing tiers. AI now sits inside that market as both a tool the attackers use and a surface they target.

The piece lines up with a sharp argument on the defence side. CSO published a clean case that organisations should stop treating AI governance as a review layer and make it release infrastructure, bolted into the deployment pipeline like CI/CD, not stapled onto the side as a quarterly check. Read together the message is uncomfortable: the attackers got industrialised while the defenders are still building governance as a slide deck.

For anyone running AI in production the implication lands directly. Your AI surface is a new market the extortion economy is pricing, and the gap between offensive industrialisation and defensive bureaucracy is exactly the window they monetise. Governance-by-review is the security posture they want you to have.

Here's what works: Treat every AI deployment as a security release milestone, not a launch milestone. Same gates as a database migration: threat model, exposure surface, recovery path, named owner. If your AI release process is faster than your security review process, you have already given the extortion economy your next case study.

7. Trump Admin Just Reframed Tech Critics As ”Terrorists”

In the political-risk lane this week, a development that should land on every comms and policy team. Reporting indicates the Trump administration is zeroing in on labelling anti-tech voices as ”terrorists” inside the broader AI regulation framing. The vocabulary shift matters: it moves AI critics from ”stakeholders” into a category the federal government has named instruments to act on.

The same week, state lawmakers continued pushing AI rules in opposite directions across jurisdictions. The pattern is clear: federal rhetoric is hardening pro-deployment, state rhetoric is fragmenting along partisan lines. For an enterprise running AI across state and federal customers, the policy ground is now uneven in ways procurement contracts written even six months ago did not contemplate. The risk is no longer ”what will the law be.” It is ”which political frame will my own AI claims be read inside next quarter.”

The harder read is operational. As the federal frame politicises, ”responsible AI” stops being a neutral term and starts being a political marker. Internal AI governance documents, external communications, sustainability claims, all of it is now legible inside a sharply polarised debate. The naming choices in your policy docs are about to matter more than they did last quarter.

Here's what works: Pull your public-facing AI statements and ask one question: ”Does this read like a marker in a culture war, or like an operating discipline?” If it sounds like the first, rewrite as the second. Operational language survives political cycles. Slogans do not.

Signal vs. Noise

🟢 Signal: Agentic deployment plumbing. The agentic-AI conversation gained sharp influence this week while raw mention volume actually dipped, the sign that buyers stopped reading ”we have agents” and started buying the secure connectivity, identity, and reserved compute that put agents into production. Snowflake's Natoma deal and DigitalBridge's power-AI combine are where the named buying authority moved. Most coverage is still treating it as a model story and missing where the money landed.

🔴 Noise: Generic ”Machine Learning”. The ”machine learning” label pulled the heaviest mention count again this week but lost real influence sharply, a clean read on a buzzword aging out of operating language. Anyone still routing AI strategy through ”ML projects” rather than agent-pipeline projects is working from a 2024 map of where buying authority actually sits today.

From the 190K

We scanned 190,000 articles this week. Here's what no one's talking about:

Snowflake reserved $6 billion of AWS capacity and bought an identity-for-agents company, DigitalBridge and ArcLight merged to own the power-AI infrastructure layer, and a hard look at enterprise rollouts found only 12 percent of AI pilots reach production, all inside the same 24-hour window.

Each desk reads these as unrelated stories. The enterprise software desk covers Snowflake. The infrastructure-finance desk writes up DigitalBridge. The consulting desk circulates the pilot-conversion number. Read them on the same morning and the picture sharpens: the AI buildout just got priced honestly. The model layer is commodity, the integration layer is where the strategics are spending, and the production gap, eight pilots in ten failing, is exactly why agentic plumbing and power-infrastructure ownership are the assets being snapped up. The strategic move on Friday is to stop budgeting AI by which model you are testing and start budgeting it by who owns the production path: identity, compute reservation, governance pipeline, and the named human who can prove a conversion rate above twelve percent.

By The Numbers

• 12% of AI pilots make it to production — Cross-industry conversion rate from a deep enterprise rollout analysis, with 46% of POCs abandoned before production and 42% of firms scrapping most AI initiatives in 2025. The honest version of ”AI ROI.”

• $6 billion AWS commitment over five years — Snowflake's capacity reservation backing its agentic-enterprise bet. When the secondary software vendor reserves nine zeros of compute, the buyers know what is coming next quarter.

• 10 million patient calls at 99.9% clinical safety — Hippocratic AI's operational milestone on DigitalOcean cloud plus NVIDIA Blackwell Ultra. The number that ”milestones over multiples” diligence is going to ask everyone else to match.

• $8 to $15M initial, $1 to $5M annual EU AI Act compliance for large orgs — The regulatory budget bar landing on high-risk AI applications August 2, 2026. Plan it like a tax, not a project line.

• Cursor: $1B annualized revenue, $29.3B valuation, 50,000 enterprises — The developer-tools layer at full scale, one signal that value is migrating into the workflows AI inhabits, not just into the models behind them.

• $250M from the OpenAI Foundation to study AI's economic impact — Quarter-billion-dollar grant aimed at the labor and economic consequences of the technology the funder also ships. The macro question just got a budget.

• See what's rising in our 190K-article corpus this week →

Deep Dive: The Venue, Not The Record

Let me take you back to a festival, the morning before doors open. The lineup is sold out. Everyone knows who is playing. But walk through the field at 10 a.m. on show day and what is actually being built is the venue, the power feeds being run, the stage rigging going up, the cooling for the gear, the security perimeter, the medical tent, the queues. The artists arrive last. They plug into wires somebody else laid the day before. This week the AI industry stopped pretending the artist was the product and started paying for the venue.

The Plumbing Got Priced

Snowflake's $6 billion AWS commitment alongside the Natoma acquisition is not a software deal, it is a venue contract. The reserved capacity buys the megawatts and racks an agentic platform actually needs to run at scale. The Natoma piece buys the access control, the SSO equivalent for software that acts on your behalf. Neither shows up on a model leaderboard. Both decide whether the model can even be plugged in next quarter, or whether it sits in a slide deck waiting for procurement to find the wiring.

The Power Wins Over The Model

DigitalBridge and ArcLight merged to be the asset manager standing at the convergence of power, AI, and digital infrastructure. The same week, a Grundfos piece said water and energy will decide whether Europe can scale AI at all, and a data-centre industry briefing raised the bar on design and environmental certification. Three signals, one read: the bottleneck moved from algorithms to amperes. The lab that wins the next year is the one whose power contract is signed.

The Production Path Is The Job

And under all of it, the operating reality: only 12 percent of AI pilots reach production. Eight in ten die in handoff, bad data, no owner, no place for the output to land. The infrastructure deals are pricing the production path. Hippocratic AI's 10-million-call milestone is the proof that path exists when you build it. The pilot-purgatory number is the proof most teams never did. The gap between those two outcomes is exactly the asset Snowflake, DigitalBridge, and ArcLight just paid up to own.

What Actually Works

1. Budget by venue, not by record. Plan AI spend by capacity reservation, identity layer, governance pipeline, and named production owner, not by which model you are benchmarking this quarter.

2. Demand operational evidence. Volume number, error rate, infrastructure named, time window stated. Hippocratic shipped one in writing. Ask for the same shape from every vertical-AI pitch.

3. Lock in capacity early. If your AI provider is buying compute on the spot market, your roadmap rides their spot market. Multi-region, multi-supplier capacity is the new procurement default.

4. Treat governance as release infrastructure. Bolt it into the pipeline like CI/CD, not as a quarterly review. Governance-by-slide-deck is what the extortion economy prices.

Anyone can buy the headliner. The night is made or broken by the people who laid the cables, signed the power contract, and set the perimeter before the gates opened. This week the smart money stopped paying for the act and started paying for the field.

What's Coming

Identity-For-Agents Becomes A Category

Snowflake's Natoma acquisition names the layer most agentic vendors are missing today. Expect a roll-up of secure-connectivity, agent-identity, and SSO-for-software companies through the back half of the year. Watch who else gets bought, that list will draw the real boundaries of the agentic stack faster than any model release will.

Capacity Becomes The Procurement Lead

DigitalBridge and ArcLight's combined positioning tells you where the supplier power is moving. Expect AI procurement reviews to start asking about power source, region, water access, and certification standards, questions that read like utility-board language because, increasingly, that is the actual constraint.

The 12% Number Becomes A Board Question

The pilot-to-production conversion gap is about to leave the consultant blog and hit the audit committee. Expect 2026 board packs to include a named conversion rate and a deadline for closing it, with the EU AI Act's August 2026 high-risk window working as the forcing function.

For Your Team

Strategic purpose: Friday is the day this week's pattern resolves into one operating change before Monday. The week named it plainly: the AI conversation moved from which model you bought to where it actually plugs in. Your edge lives in who owns the venue, the capacity, the identity layer, the production path, and naming those owners is the work most strategy decks still skip.

Friday's meeting prompt: ”If 12 of every 100 AI pilots reach production, and the smartest money this week paid $6 billion for plumbing and merged a firm to own the power layer, then which 12 percent of our AI portfolio is actually plugged in, and who is named on each one? Or are we still optimising the model when the wiring is the constraint?”

The Plug-In Discipline Framework:

1. Name the venue owner — For every AI initiative, identify the person accountable for capacity, identity, and the production path. No named owner, no green light.

2. Budget the wiring before the act — Allocate the integration, governance, and capacity-reservation spend before approving the model spend. The cheap model is what stalls without paid plumbing.

3. Demand the milestone shape — Volume number, error rate, infrastructure named, time window stated. From every vendor, every pilot, every internal claim.

4. Set a conversion target — Decide what percentage of pilots must reach production in 2026, and stop greenlighting new ones until current ones close that gap.

5. Bolt governance to release — Treat AI governance as deployment infrastructure on the build pipeline, not a quarterly review. The extortion economy industrialised while compliance stayed a slide deck.

Share-worthy stat: Only 12% of enterprise AI pilots make it to production, while Snowflake just reserved $6 billion of AWS capacity to underwrite the agentic enterprise plumbing, the gap between AI demos and AI deployments, priced in the same week. Drop that on the next strategy review and the ”are we buying the model or the venue” conversation writes itself.

Go deeper: Track where AI capacity and integration are actually landing in real-time →

The Track of the Day

”Multiples still matter, but AI tends to put more weight on execution evidence. The more 'in-flight' the value, the more the conversation shifts from a single price number to a set of measurable milestones.”
— From a piece on how AI is changing valuation, the milestones-over-multiples shift

Today's set closes on the record everyone in diligence is now playing: the receipt track. The whole week was the market refusing to pay for promised outcomes and starting to pay only for measured ones, calls completed, capacity reserved, pilots converted, security gates passed. The operator running on demo reels is mixing blind. The one walking into Friday with a number for every claim is the one whose set actually lands.

Yves Mulkers, your data DJ, mixing 190,000 articles into the tracks that actually matter.

We scanned 190,000 articles this week so you don't have to. Data Pains → Business Gains.

Published: May 28, 2026 | Curated by Yves Mulkers @ Ins7ghts