| The Agentic Enterprise |
AK · Tue, Jul 14, 2026 · 7 min |
|
|
Tuesday, July 14, 2026
Stop renting your intelligence.
Satya Nadella and Alex Karp, rarely allies, are making the same argument: the AI you rent is learning your business, and you are paying for the privilege.
Nadella's new essay says you pay for intelligence twice, once in money and again in the proprietary knowledge you must reveal to make it useful. Karp says enterprises are burning cash on tokens and handing their IP to OpenAI and Anthropic. The uncomfortable part is that both men are also selling the cure.
|
|
The Lead
On July 13, Satya Nadella published the sharpest thing a hyperscaler CEO has said about the AI trade all year.
He called it "The Reverse Information Paradox." The core line: "You essentially pay for intelligence twice, once with money, and again with something even more valuable: the proprietary knowledge you must reveal to make that intelligence useful." Two weeks earlier, Palantir's Alex Karp had spent twenty minutes on CNBC making the cruder version of the same point, that enterprises are burning money on tokens and handing their IP to OpenAI and Anthropic in the process.
When the CEOs of Microsoft and Palantir independently arrive at the same warning, the question stops being whether to use frontier models and becomes who keeps what they learn from you.
|
|
Karp and Nadella agree: you are paying for AI twice.
| N |
adella's July 13 essay reframed the entire enterprise AI purchase in one sentence. You pay once in dollars for tokens, he argued, and again in something more valuable: the proprietary knowledge you must feed a model to make it useful. Models learn from "exhaust," the prompts, the tool calls, and above all the corrections users make when the model is wrong, and every correction is distilled into capability that does not stay with you. "In consuming intelligence, you are creating intelligence," he wrote. "And what you create should belong to you." |
Two weeks earlier, on July 1, Palantir's Alex Karp made the blunt version on CNBC, declaring that "something has gone completely wrong" with the token business and that enterprises get little value while risking their IP to OpenAI and Anthropic. He was pitching Palantir's new Nvidia partnership for "secure AI," and David Sacks backed him by pointing to the Anthropic-Figma episode, where a design collaboration reportedly grew into a competing product.
|
In consuming intelligence, you are creating intelligence, and what you create should belong to you.
|
For enterprise leaders, this reframes the central AI question. It is no longer only "which model," a race that has largely converged on capability. It is "who retains the learning," which is a data-governance and contract question your legal and security teams should own, not a benchmark your ML team settles. Nadella's prescription is concrete: keep ownership of your data and feedback, build proprietary learning environments in your own cloud, and put an orchestration layer between you and any single model so you can switch. The uncomfortable footnote is that this is not literally true the way Karp implies. The leading labs do not train on enterprise prompts by default, and offer contractual opt-outs. The real exposure is narrower, the "design partner" category, where a vendor sees enough of your workflow to build the thing you were going to buy.
|
The Spearhead Take
Both men are talking their book, and both are still right. Palantir sells the "secure AI" alternative; Microsoft sells the cloud and orchestration layer you would build sovereignty on, which is why analysts called Nadella's framing smart and self-serving in the same breath. Discount the sales pitch, keep the insight. Make data governance a first-order procurement term: get no-train and no-retention guarantees in writing, own your evaluation sets and the corrections your people generate, and never let a design-partner relationship expose more of your workflow than the contract protects. Sovereignty is not self-hosting everything, which most firms cannot do at frontier quality. It is refusing to give away the one asset a rented model cannot replace, which is what your business specifically knows.
|
|
|
The Obvious & The Overlooked
The warning is the headline. That both men sell the cure is the story.
|
The Obvious
Nadella and Karp both warned that enterprises risk handing proprietary knowledge to the frontier labs.
The IP-leakage worry has gone mainstream at the CEO level. TechCrunch
Nadella's answer is "sovereignty": own your data, build proprietary learning environments, keep an orchestration layer.
Own your intelligence, do not rent it query by query. 24/7 Wall St.
Karp used the argument to pitch Palantir's Nvidia partnership for secure AI.
The warning doubles as a sales motion. CNBC
|
The Overlooked
The leading labs do not train on enterprise prompts by default and offer opt-outs, so the literal IP-theft claim is overstated.
The scary version is mostly wrong; the structural version is not. Fortune
The real exposure is the "design partner" relationship, where a lab sees enough of your workflow to build a competitor.
The Anthropic-Figma episode is the cautionary case, not the average API call. Benzinga
Both men are selling the cure they diagnose, which analysts flagged immediately.
Nadella's sovereignty runs on Azure; Karp's secure AI runs on Palantir. Computerworld
The asset that stays yours is the correction data, the feedback your people generate fixing the model's mistakes.
Own that pipeline and you keep the compounding value even while renting the model. Explainx
|
|
|
Moving Pieces
Five developments worth a CIO's attention.
Deals
Together AI raises $800M as open-model demand triples
Together AI closed an $800 million Series C on July 1 at an $8.3 billion valuation, more than double its mark from sixteen months ago, led by Aramco Ventures with NVIDIA, Vista Equity Partners, and General Catalyst in the round. The company runs open-source models as a service, and its bookings crossed $1.15 billion last quarter as open-model usage across the industry tripled in twelve months. The enterprise read: if Nadella's sovereignty pitch has a shopping list, self-hostable open weights are on it, and this is the infrastructure making that practical. A neocloud growing this fast on open models is the clearest sign that "own it" is becoming a real option, not just a keynote line, for teams that want the model inside their own perimeter.
Infrastructure
Anthropic chases OpenAI into custom silicon, via Samsung
Anthropic is in early talks with Samsung to build a custom AI chip on a 2-nanometer process, aimed at inference rather than training, according to reporting first surfaced by The Information. The project is preliminary, with functions and performance targets still undefined, and it follows OpenAI's Jalapeno inference chip, built with Broadcom and unveiled last month. Broadcom's CEO has cited roughly 50% inference cost savings versus standard GPUs in early testing. The enterprise read: Anthropic is the follower here, not the leader, and every frontier lab is now vertically integrating down to the silicon to escape the GPU tax. That is good for the eventual price of inference you buy, and a reminder that your model vendor's margins, and therefore its pricing, are still a moving target.
Deployment
The Palantir playbook becomes everyone's playbook
AWS stood up a $1 billion Forward Deployed Engineering organization that embeds pods of five or six engineers, paired with the agents they build, inside customers on 45-day sprints, with the Allen Institute, Cox Automotive, the NBA, the NFL, Ricoh, and Southwest Airlines among the first. It joins OpenAI's and Anthropic's own deployment arms, both backed by private equity. The forward-deployed model, long Palantir's signature, is now the whole industry's answer to a hard truth: enterprises fail at AI not because models are weak but because integrating them into decades of messy workflow is brutal. The enterprise read: the vendors have concluded that selling you a model is not enough, so they are selling you the humans to make it work. Useful, but budget for the day the pod leaves.
Markets
The AI IPO race is now real, with two filers
The back half of 2026 has its first genuine market test. OpenAI filed a confidential S-1 on June 8, targeting a September debut at $730 billion to $850 billion, days after Anthropic filed its own draft on June 1. The two enter from opposite positions: OpenAI leads on consumer reach and does not project profitability until 2030, while Anthropic is the revenue leader at a roughly $47 billion run rate and, by its own projection, the first frontier lab to post a quarterly operating profit, in the current quarter. The enterprise read: your two most important model vendors are about to have public balance sheets and quarterly earnings calls. Pricing, roadmap, and support decisions will start bending to shareholder expectations, so read the S-1 risk factors before you sign a multi-year commit.
Security
As agents get platform access, securing them becomes a line item
The same access that lets agents reach enterprise tools also lets them reach enterprise damage. Researchers this month documented a ransomware campaign run almost entirely by an autonomous AI agent, which exploited a Langflow vulnerability with little human direction, and a new crop of vendors is racing to sell the countermeasure. Straiker raised $64 million in Series A funding, backed by Citi Ventures and Workday Ventures, to discover, test, and protect enterprise AI agents. The enterprise read: agent security is becoming its own budget category, separate from the model and the platform. If you are giving an agent credentials and autonomy this quarter, fund the red-teaming and runtime controls in the same breath, because that agent is an insider threat you provisioned yourself.
|
|
On the Radar
Eight signals, sharpened.
| Workforce |
Anthropic hired Monzo co-founder Tom Blomfield onto its compute team. The talent war now pulls fintech operators into frontier-lab infrastructure roles, not just researchers. Tech Startups |
| Deals |
CarbonSix raised $40 million in Series A funding for physical AI in manufacturing. Robotic intelligence for the factory floor is drawing real capital, not just chat-window demos. Crescendo AI |
| Deployment |
Experity acquired Exdion Healthcare to automate coding, billing, and compliance for urgent-care operators. Vertical AI roll-ups are targeting the unglamorous revenue-cycle work where ROI is measurable. Daily AI Brief |
| Research |
Gartner projects 40% of enterprise apps will embed task-specific AI agents this year, up from under 5% in 2025. The embedding is happening inside software you already own, not just in standalone tools. Gartner |
| Product |
Google's Gemini Enterprise platform now offers first-class access to 200-plus models through Model Garden, including Gemini 3.1 Pro and open models like Gemma 4. Model choice, and the freedom to switch, is becoming a platform feature. Google Cloud |
| Markets |
US venture funding hit a record $412.7 billion in the first half of 2026, with AI deals dominating. Capital is not the constraint on enterprise AI; deployment is. SiliconANGLE |
| Deals |
LeapXpert closed $180 million in growth financing for compliant enterprise communications. Governance and compliance tooling is where regulated industries are spending their AI budget. Crunchbase News |
| Deals |
Brand Engagement Network completed its acquisition of Munich-based enterprise software vendor Cataneo. Consolidation is reaching the mid-market and beyond the US, not only the megacaps. Tech Startups |
|
|
Quick Hits
The wider field, one line each.
| KredosAi raised a $7 million Series A led by BMW i Ventures for AI-driven debt collections. The AI Insider |
| Glean shipped its Enterprise Agent Development Lifecycle to help companies build, govern, and measure agents. Glean |
| Glean added support for NVIDIA Nemotron 3 Ultra, expanding cheaper open-model options. Glean |
| Together AI's annual bookings crossed $1.15 billion as open-model usage tripled in a year. BusinessWire |
| Palantir and NVIDIA partnered on a secure enterprise AI stack that Karp pitched as the alternative to renting tokens. CNBC |
| OpenAI's custom inference chip, Jalapeno, built with Broadcom, debuted last month. The Next Web |
| Anthropic signed a roughly $19 billion data-center lease with sustainable-computing firm TeraWulf. CNBC |
| Gartner projects over 40% of agentic AI projects will be canceled by 2027 on runaway cost and unclear ROI. Gartner |
| IDC found 88% of AI proofs-of-concept never reach wide-scale deployment. Beri.net |
| Google opened a $750 million innovation fund for partners building and deploying AI agents. Google Cloud |
|
|
The Number
86%
Of every US venture dollar, first half
The share of US venture funding that went to AI companies in the first half of 2026, about $356 billion of a record $412.7 billion raised.
Capital has picked its winner, and it is the category, not any one company. For an enterprise buyer, the read is not that AI is hot, you knew that. It is that nearly every vendor on your shortlist is now flush, aggressive, and racing to spend, which means faster roadmaps and louder sales motions, but also less pricing discipline and more vendors that will not exist in their current form by 2028.
|
|
Counter-Signal
Risk
Sovereignty is a luxury most enterprises cannot afford yet.
Nadella and Karp make owning your intelligence sound like a choice every company should make. For most, it is not one they can make this year. Building a proprietary learning environment assumes you have clean data, an ML team, an orchestration layer, and agents that actually reach production, and the data says most do not. Gartner puts the failure rate of AI agent pilots at 89%, IDC finds 88% of AI proofs-of-concept never scale, and Gartner expects more than 40% of agentic projects to be canceled by 2027 on cost and unclear ROI.
Hold both truths. Sovereignty is the right long-term posture, and giving away your correction data is a real cost. But for the 89% still stuck between pilot and production, the frontier labs' rented intelligence is the only intelligence that works right now, and refusing it on principle means shipping nothing while competitors ship something. The pragmatic path is not to reject rented models, it is to rent them on terms that protect what matters, contractual no-train guarantees and ownership of your evals and feedback, while you build the muscle to own more over time. Sovereignty is a destination, not a switch you flip because two CEOs told you to.
|
|
From the Field
The two men are an unlikely pair, and this month they said the same thing from opposite ends of the industry: the intelligence you create by using AI should belong to you, and right now much of it does not.
They are both talking their book, and it is worth saying plainly. Palantir sells the secure alternative. Microsoft sells the cloud and the orchestration layer you would build sovereignty on. When two vendors describe a disease and each happens to sell the cure, a careful buyer discounts the urgency and keeps the diagnosis. The diagnosis here is sound. The value in enterprise AI is migrating from the model, which is converging and getting cheaper, to the accumulated, specific knowledge of how your business actually runs, and that knowledge leaks one correction at a time when you never thought to keep it.
So do the unglamorous thing. Read your data-processing terms before your next renewal and get the no-train, no-retention language in writing. Own your evaluation sets and the feedback your people generate, because that is the asset a rented model cannot reproduce. Keep an abstraction layer between you and any one vendor so switching stays cheap. You do not have to self-host the frontier to refuse to give it away. Sovereignty, in the only version most companies can afford this year, is just remembering that what you teach the machine is worth keeping.
Let's get to production,
AK
|
|
|
|
Anthropic is a Spearhead technology partner, and its Claude model produced this edition under human editorial direction. This edition applies more scrutiny to Anthropic, not less: the Big Story reports Alex Karp's and, structurally, Satya Nadella's warnings that enterprises risk their IP to OpenAI and Anthropic, cites the Anthropic-Figma design-partner episode as the cautionary case, and names Anthropic elsewhere as a follower chasing OpenAI into custom silicon and as one of two IPO filers. The same critical frame is applied to OpenAI throughout, and the edition includes Fortune's counterpoint that the literal IP-theft claim is overstated because the leading labs do not train on enterprise prompts by default and offer opt-outs. No favorable framing was applied on account of the partnership. Nadella's July 13 essay "The Reverse Information Paradox" and its quotes are reported by TechCrunch, 24/7 Wall St., Computerworld, and Explainx; the original was p | |
