Kuaishou's Kling AI Hits $90M, Oncology's 22-Point Gender AI Divide

So, What Actually Happened?

So, Friday. The headliners packed up, the venue is dark, and what's left on stage is the cleanup crew nobody photographed. We scanned 190,000 articles this week so you don't have to. A fresh oncology survey out of MD Anderson found 93% of practitioners want AI training most of them never got. Kuaishou's Kling AI booked RMB650M in a single quarter at 300%+ growth, a real-money AI video business hiding inside a Chinese earnings call. A new SSRN paper named the Deployer Multiplier, the legal mechanic that pushes EU AI Act risk down from the lab to whoever plugs the model in. And on the security side, SentinelOne turned sovereign AI into a regional strategy across KSA and UAE.

The Bottom Line: Last week was about who builds the venue. This week is about who staffs the door. The training, the deployer, the cost, the auditor. The cleanup crew nobody photographed just became the most expensive line item in the AI budget.

The Tracks That Matter

1. 93% Of Oncologists Want AI Training Most Never Got

Here's the receipt for ”AI in healthcare” diligence. A fresh MD Anderson survey found 93% of oncology practitioners want dedicated AI training for research or clinical care, while 58% already use AI several times monthly to daily and 15% have never touched it at all. The training they want has nowhere to come from. Most institutions ran the pilot, skipped the curriculum, and shipped the tool to staff who learned on the job, badly.

The same survey landed the demographic split nobody is comfortable naming out loud. 70% of male respondents are frequent AI users versus 48% of women, with age 60+ also showing significantly lower odds of frequent use. Same hospital, same access, same tools, twenty-two-point gap in actual workflow adoption. Read it next to a separate piece this week on AI-driven precision learning in neurosurgical residency, which found a structured AI-enhanced curriculum cut entrustment time and reduced inter-resident variability. Training isn't theoretical anymore. It's the operating difference between an ”AI productivity uplift” projection that lands and one that quietly dies inside the building.

The boardroom version of this is unfun. Most healthcare AI diligence still asks about model accuracy, vendor security, and FDA path. None of those questions surface a 22-point gender adoption gap inside the same workforce, the kind of variance that nukes any productivity uplift number a vendor put in the deck. The thing killing healthcare AI ROI isn't model quality. It's that your own people split into a fast lane and a slow lane based on training nobody scheduled.

Here's what works: Pull AI-tool usage data by department and demographic cohort before the next vendor review. If the spread is wider than 15 percentage points between any two cohorts, the gap isn't the tool, it's the missing training scaffolding. Allocate the curriculum budget before the license budget, not after.

2. China's Kling AI Just Booked $90M In One Quarter

Hidden in a Chinese earnings call this week, a number Western AI coverage mostly skipped. Kuaishou's Q1 disclosure landed Kling AI revenue at over RMB650 million, a 300%+ year-over-year jump for the video generation model Kuaishou now treats as its marquee product. The overall company revenue hit RMB33.7 billion for the quarter, up 3.4% YoY, with the Kling line carrying the headline growth.

Read it against what most US tech wires keep arguing about. The debate about whether generative video has a paying market is still live in English; Kling just put roughly $90 million USD on the board in three months, with the AI segment growing roughly a hundred times faster than the parent business. The other Kuaishou disclosures fill in the operating shape: gross margin compressed to 51.2% from 54.6%, cost of revenue up 11.1% on higher bandwidth and revenue-share. Exactly what scaling a generative video model into actual production looks like on the income statement. Real revenue, real margin pressure, real category.

For anyone routing their AI strategy through ”is generative video commercial yet,” this is the receipt. Paying customers for AI video exist right now at meaningful scale, they're just mostly in China and mostly invisible to the English-language analyst desk. If your media, advertising, or content roadmap is waiting for a Western incumbent to validate the category before you start benching tools, you're trailing a market that already cleared $360M annualized on a single product line at a single platform.

Here's what works: Add a line to your 2026 content roadmap naming which Chinese-stack AI video tool you've evaluated and benched against Runway or Sora. Skipping the comparison because the vendor isn't on a Western leaderboard is the cleanest way to ship a 2024 budget for a 2026 market.

Try It Yourself

3. The ”Deployer Multiplier” Just Cracked AI Liability Wide Open

A fresh SSRN paper this week named the legal mechanic most enterprise AI counsel has not yet priced in. Annex A of the EU AI Act readiness research, titled ”The Deployer Multiplier,” sets out how high-risk obligations cascade from the model lab down to every organization that deploys the model in a regulated context. The compliance bar no longer stops at OpenAI, Anthropic, or Mistral. It lands on the hospital, the bank, the insurer, the school district.

The shape matters. Until this paper, most internal AI governance setups scoped the EU AI Act as a vendor problem. The Deployer Multiplier reframes it as an operations problem. Articles 25, 26, and 27 each carry independent enforcement triggers on deployers, not just developers, and the Article 27 FRIA (Fundamental Rights Impact Assessment) requirement reads like a SOX-style internal control regime grafted onto AI deployment. The Article 25 reverse-bascule mechanic means a ”general-purpose” model becomes a ”high-risk system” the moment your own team configures it for a regulated use case. The procurement contract that said ”vendor handles AI Act compliance” needs a re-read this quarter.

The August 2 high-risk window opens in nine weeks. Most enterprise AI governance is still scoped to vendor selection and security review, not deployer-side documentation and post-market monitoring. The in-house team that connected the model to the clinical workflow, the credit-decision flow, or the hiring pipeline now carries obligations the lab does not carry for them. That is the multiplier, and it lands on the people who pressed ”deploy.”

Here's what works: Get legal and the AI deployment leads in the same room before mid-July. List every production AI use case, tag each as general-purpose model, high-risk deployment, or both. Anywhere the answer is ”both,” assume the Deployer Multiplier applies and start the FRIA paperwork now, not in September.

4. Tempus Wired An Agentic Co-Pilot Into Physician Workflows

In a week mostly about plumbing, here's a vertical deployment with actual shape. Tempus AI shipped an agentic GenAI co-pilot into its Hub platform for oncology physicians, pairing the agentic reasoning layer with Tempus' existing structured patient data backbone. Not a chat widget bolted onto a portal. An agent that sits inside the clinical workflow where the oncologist already lives, reading and writing into the record under named identity and audit.

The shape connects two things last week's news left dangling. The agentic infrastructure plays bought the layer that says ”agents can act on your behalf without breaking SSO.” Tempus is the first vertical proof a real regulated deployment is now landing on top of that layer. The piece reads next to a Cognizant-Travelport AI deployment partnership signed the same week for travel-industry operations. Different sectors, same shape: structured data backbone, agent inside the workflow, identity and audit from day one, named human as the accountable approver.

For healthcare specifically this is the version of ”AI in the EHR” that survives a clinical safety review. Tempus' assembly is the operating shape regulators reward and insurers price, the bolted-on diagnostic chatbot pattern that won 2024 headlines is not where 2026 deployment dollars are landing.

Here's what works: If you're scoping a vertical AI deployment, copy the Tempus assembly: structured data backbone you already own, agent layer inside the workflow your team already uses, identity and audit baked in before the first user touches it, named human as the accountable approver. The standalone chatbot is the pattern losing the procurement cycle.

5. Ping Identity Just Made Login Work For Software That Acts

Here's the layer most agentic vendors keep hand-waving. Ping Identity added native controls for AI agents in business deployments this week, treating agent identity as a first-class category instead of a feature add. SSO, lifecycle, consent, and audit for software that acts on your behalf, not just for humans.

Read it against last week. Snowflake bought Natoma for almost exactly this problem; now the established identity stack is shipping native agent controls on the same week. Two signals, same window: the agent-identity layer is consolidating from ”specialty startup” into ”category every identity vendor must support.” The IdP vendors treating it as an upsell are about to lose enterprise procurement to the ones treating it as core strategy. The procurement template that does not yet have an ”agent identity provider” row is already late.

For your AI vendor reviews the consequence is direct. Any agentic AI pitch this quarter without a clear answer for ”who handles agent identity, lifecycle, and consent” should not get the second meeting. The vendors whose answer is ”we partner with Ping, Okta, or your existing IdP” are pitching a stack that actually runs in your environment without a separate audit committee fight.

Here's what works: Add an ”agent identity provider” line to your AI vendor scorecard, sitting alongside ”model provider” and ”compute provider.” If the agentic vendor cannot name a real identity stack they integrate with, the procurement gate stays closed until they can. The SSO-for-agents check happens before the contract goes to legal, not after.

6. Portal26 Shipped The First Cost Throttle For AI Agents

Buried in the product wires this week, the FinOps problem every agentic deployment hits in week four. Portal26 launched what it calls industry-first AI agentic cost controls, naming the issue most enterprise buyers found the hard way: agents loop, retry, escalate, and rack up token bills nobody on the team owns. The throttle treats inference cost the way FinOps treated cloud spend a decade ago.

The shape is overdue. Anyone running agentic AI in production this year has watched the API bill spike on a Tuesday morning because a single misconfigured agent decided to retry a tool call 600 times before giving up. Cost guardrails for human users (rate limits, quota caps) do not map to agent traffic; agents do not respect the social signals humans do. Portal26 is the early mover; agent-level FinOps will likely be table stakes by Q3.

For CIOs the read is that the AI budget shape is about to flip. Model spend is going up because agents are calling more often, not because per-token prices are rising. Without a cost throttle, the cloud-cost lesson of 2014 (visibility before optimization) plays out again in 2026 with bigger numbers. The vendors that ship cost controls first will price themselves as platforms, not features.

Here's what works: Add agent-level cost telemetry to your AI deployment checklist this quarter. If you cannot see which agent ran how many tool calls against which model in which workflow, you cannot bound the bill. Treat agentic FinOps as a 2026 budget item, not a 2027 cleanup project.

7. SentinelOne Just Turned ”Sovereign AI” Into A Security Strategy

In the geopolitical-AI lane this week, a deployment naming where the real friction is moving. SentinelOne extended its sovereign AI-driven cybersecurity strategy across KSA and UAE, framing sovereign AI not as a slogan but as a security architecture: in-region model hosting, in-region data, in-region threat intel, all under the local regulator's jurisdiction. The Gulf is the proving ground, the play is global.

It lines up with two other geographic-AI signals this week. Statista's data shows AI talent concentration shifting toward India and away from Silicon Valley's monoculture, and China just released its 2026 plan for building an IP powerhouse naming AI assets as a national priority. Three signals, one read: the working assumption that AI infrastructure is a US-centric monoculture is rapidly dating itself.

The enterprise consequence is plain. Your AI vendor selection criteria are about to include data residency, model residency, and regulatory jurisdiction in ways your 2024 procurement template does not contemplate. The customer who needs to operate in Saudi, the EU, and India simultaneously is now buying three AI stacks, not one, and ”sovereign-by-default” architecture is the only way to ship without lawyering each integration separately.

Here's what works: Map every AI deployment to the regulators that touch it. Anywhere a single deployment crosses two sovereign jurisdictions, treat that as architectural debt, not a compliance footnote. Vendors with sovereign builds already running will price their advantage in 2027; the ones bolting it on after the fact will price their delay.

Signal vs. Noise

🟢 Signal: Agentic AI deployment. Agentic AI's real influence rose sharply on Thursday across the wires as named vendors (Tempus, Cognizant-Travelport, Ping Identity, Portal26) shipped production deployments instead of pilots. Most coverage is still writing about ”AI agents” as a future-tense product category and missing where the actual procurement signatures are landing this quarter.

🔴 Noise: Generic ”Machine Learning”. The ”Machine Learning” label pulled heavy volume again on Thursday but kept losing ground in operating language, exactly the shape of a buzzword aging out of buyer vocabulary. Anyone still routing the AI strategy review through ”ML projects” rather than agentic deployment 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:

MD Anderson found 93% of oncology practitioners want AI training they don't have, Tempus shipped an agentic co-pilot directly into the oncology workflow, and Portal26 launched the first agent-level cost controls, all inside the same 48-hour window.

Each desk reads these as unrelated stories. The healthcare research desk covers the ASCO survey. The medtech equity desk writes up Tempus. The DevOps press picks up Portal26. Read them on the same morning and the picture sharpens: the AI deployment layer has shipped past ”does it work” and into the hard operating questions. Who is trained, who is named, who is metered, who pays. The strategic move before Monday is naming the human side of every AI initiative before the model side, because the model is no longer the limiting reagent. The training scaffolding, the cost telemetry, and the named accountable owner are.

By The Numbers

• 93% of oncology practitioners want dedicated AI training — MD Anderson Cancer Center survey, with 58% already using AI several times monthly to daily and 15% never using it. The training scaffolding is the gap, not the tool.

• RMB650 million Kling AI revenue in one quarter, +300% YoY — Roughly $90 million USD of paid AI video generation, in a single Chinese quarter on a single platform. The category Western press still debates is already monetizing offshore at scale.

• 70% of men vs 48% of women are frequent AI users in oncology — Twenty-two percentage point gender adoption gap inside the same workforce. Any ”AI productivity uplift” projection that doesn't decompose by cohort is fiction.

• RMB33.7 billion Kuaishou Q1 revenue, +3.4% YoY — The base business grew 3.4 percent; the AI line grew more than 300. The composition shift is the actual story, not the topline.

• $2 million enterprise loss from one isolated AI test gone live — A single deployment moved from sandboxed test to corporate network with the exact same code, took the data with it. The cost of skipping integration testing got named in cash terms.

• BCG: 30 to 50% process acceleration with controlled agentic AI — Upside is real, but the precondition is governance from day one. Firms hitting the upper end built the rails before flipping the switch.

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

Deep Dive: The Operating Floor

Let me take you back to a club at 2 a.m. The headliner finished an hour ago, the venue is still loud, and what makes or breaks the night now is the operating floor. The bartenders pouring fast, the security checking IDs, the runner restocking ice, the manager watching the till. Nobody photographs the operating floor. It's why most clubs that throw great parties go broke. Last week the AI industry built the venue. This week the operating floor showed up uninvited.

The Training Floor

The ASCO survey put a number on the people problem most healthcare AI decks skip: 93% of oncology practitioners want AI training, only 58% use AI regularly, and women trail men by 22 points in actual frequent use inside the same hospitals. The gap isn't a tool problem. It's a curriculum problem. A separate neurosurgical residency paper this week showed that structured AI training shaved entrustment time and reduced inter-resident variability. Proof that the gap closes when somebody actually teaches the work. Most healthcare AI budgets ship the license, skip the curriculum, and call it deployment.

The Compliance Floor

A new SSRN paper named the Deployer Multiplier, the EU AI Act mechanic that cascades high-risk obligations from the model lab to every organization that deploys the model in a regulated context. The August 2 window opens in nine weeks. The internal counsel who treated the AI Act as a frontier-lab problem is about to discover that Article 27's FRIA requirement landed on her own procurement team. The compliance floor used to sit on the lab. Now it sits on you, on the team that pressed deploy.

The Cost Floor

Portal26 launched what it calls the first agent-level cost-control product. Agents don't respect human rate-limit signals. They loop, retry, escalate, and rack up token bills with no social brake. Cost throttles for agents are about to be table stakes the way cloud-cost FinOps became table stakes a decade ago. The CIOs who already wrote a 2027 line item for it are ahead. The rest are about to discover their AI bill on a Tuesday morning Slack thread, the same way they once discovered AWS bills in 2014.

What Actually Works

1. Budget the training before the license: If you cannot name the person who teaches the workflow, do not buy the workflow tool. The 22-point gender adoption gap in the ASCO data is the cost of skipping this.

2. Map every production AI to a Deployer Multiplier audit: Tag each use case as general-purpose, high-risk, or both. Anywhere the answer is ”both,” the FRIA paperwork starts now, not in September.

3. Meter the agent, not just the model: Add agent-level cost telemetry to every deployment. Tool calls per agent per workflow per day, surfaced to the team that actually pays the bill.

4. Name the identity layer in every procurement: If your agentic vendor's answer for agent SSO is hand-waving, the procurement gate stays closed. Ping, Okta, or your existing IdP, named in the contract before legal sees it.

The headliner walked off. The crowd is still here. What separates the venue that survives Q3 from the one that doesn't is whether the operating floor (training, compliance, cost, identity) was budgeted, named, and metered before the lights came up. The venue without the floor is just a room that loses money loudly.

What's Coming

The FRIA Backlog Hits Internal Counsel

The Deployer Multiplier paper puts an enforcement clock on Article 27 FRIA documentation that the EU AI Act community has mostly assigned to model labs. Expect Q3 enterprise legal calendars to fill with FRIA assessments on internal AI deployments, with the first significant fines landing on deployers, not developers, by year-end.

Agent-Level FinOps Becomes A Procurement Line

Portal26's industry-first cost controls names a problem every team running agents in production has already met. Expect the major cloud-cost platforms to ship native agent-cost telemetry by Q4, with finance teams demanding visibility before greenlighting new agent rollouts.

Sovereign AI Architecture Hits Western Procurement Templates

SentinelOne's sovereign AI push in the Gulf is the leading edge. Expect Western Fortune 500 procurement templates to include data-and-model residency clauses by Q1 2027 as customers with multi-jurisdiction footprints stop accepting ”we run in US-East” as a deployment story.

For Your Team

Strategic purpose: Monday is the day this week's pattern lands on the leadership team. Last week was the venue. This week is the operating floor: training, compliance, cost, identity, the lines most AI strategy decks still skip. Your edge is naming the operating-floor owner before the next pilot is greenlit, because that owner is the one preventing the 22-point adoption gap, the FRIA fine, the runaway agent bill, and the failed audit.

Monday's meeting prompt: ”If 93 percent of our practitioners want AI training they don't have, our agent bill could spike on a Tuesday morning without warning, and the EU AI Act's August 2 window lands on our deployment team and not just our vendors, then who on this team owns the operating floor (training, compliance, cost, identity) for each production AI use case? Or are we still running pilots assuming the floor takes care of itself?”

The Operating Floor Framework:

1. Name the training owner: For every AI tool in production, identify the person who runs the curriculum and tracks usage by cohort. No named training owner, no green light.

2. Run the Deployer Multiplier audit: Every production AI use case tagged as general-purpose, high-risk, or both. ”Both” starts FRIA paperwork this quarter, not next.

3. Meter the agent, surface the bill: Agent-level cost telemetry on every deployment, visible to the finance lead before the next budget cycle, not after.

4. Name the identity layer: Every agentic vendor names a real IdP integration (Ping, Okta, or your existing stack) in the contract before signature.

5. Map the sovereign jurisdictions: Any AI deployment crossing two regulators is architectural debt. Price the cost of the bolted-on sovereign build before you sign a multi-region rollout.

Share-worthy stat: 93 percent of oncology practitioners want AI training they don't have, while women in the same workforce trail men by 22 percentage points in actual frequent use of the tools their employer already bought. The adoption gap inside the same building is the cost of skipping the curriculum.

Go deeper: Track where AI deployment, training, and compliance are landing in real-time →

The Track of the Day

”These findings highlight the need for structured, accessible training and institutional and societal guidance to promote appropriate, equitable, and high-quality integration of AI into oncology research and clinical care.”

From the MD Anderson Cancer Center AI knowledge and use survey

Today's set closes on the record nobody on the AI tour is playing yet: the training track. The headliners shipped the model, the venue paid for the rigging, and the operating floor is the line that still has no name on it. The team that names the training, the compliance, the cost, and the identity before Monday's standup is the one whose set actually keeps the crowd inside.

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 29, 2026 | Curated by Yves Mulkers @ Ins7ghts