Oasis Security Bags $120M as AI Security Apocalypse Looms

Your weekly signal boost from 190,000+ articles, served with a DJ's ear for what actually matters.

So, What Actually Happened?

We scanned 190,000 articles this week so you don't have to. The pattern that jumped from the data? The AI industry is diverging. On one side, GMI Cloud just announced a $12 billion sovereign AI infrastructure complex in Japan, one gigawatt of power for a single AI campus. On the other side, First Round Capital co-founder Howard Morgan told the Economic Times that AI startup valuations are ”overheated” and the ”buy high, sell higher” playbook only works in a bubble. Meanwhile, Oasis Security raised $120 million to protect AI agents from the security threats nobody planned for, and the White House is scrambling to roll out an AI framework by Friday.

The Bottom Line: The biggest infrastructure bets in AI history are happening at the same moment the smartest investors are calling the valuations a bubble. Both can be right. Infrastructure retains value even when the companies built on it don't. The question for your organization: are you investing in the rails or the railroad companies?

The Tracks That Matter

1. $120 Million Says the Biggest AI Security Threat Is the One Nobody Planned For

When Sequoia Capital, Cyberstarts, and Craft Ventures collectively put $120 million into a single security company, it is worth asking what they see that the rest of the market does not. Oasis Security raised a Series B to build what they call ”agentic access management,” and the framing tells you everything. This is not traditional cybersecurity. This is security built specifically for a world where AI agents act autonomously inside your enterprise systems.

The timing is not accidental. The same week, Menlo Ventures published an analysis arguing that offensive AI has reached a tipping point: the tools attackers use to probe, exploit, and infiltrate are now AI-native. The old playbook of perimeter defense and access controls was designed for humans clicking through interfaces. AI agents do not click. They call APIs, chain actions across systems, and operate at speeds no human security team can monitor in real time. Governing tens of thousands of AI agents requires policy chaining, a concept most enterprises have not even started thinking about.

What makes this funding round structurally important is the category it creates. ”Agentic access management” is not a feature bolted onto existing security products. It is an entirely new layer of the security stack, purpose-built for an architecture that did not exist two years ago. The organizations deploying AI agents at scale, and there are more of them every week, need security that understands agent behavior, not just user behavior.

Here's what works: If your organization is deploying or planning to deploy AI agents, add ”agentic security posture” to your risk assessment immediately. The legacy access management tools your team relies on were built for human users. Ask your security team: do we know what every AI agent in our environment has access to? If the answer takes more than five minutes to produce, you have a gap that Oasis Security just raised $120 million to fill.

2. $12 Billion for a Data Center You Cannot Build in America

GMI Cloud announced a $12 billion, one-gigawatt sovereign AI infrastructure initiative in Kagoshima, Japan, and the location tells you as much as the price tag. Japan. Not Texas. Not Virginia. Not any of the U.S. data center corridors where every hyperscaler is fighting for the same constrained power grid. GMI Cloud and its partners, Wistron, Kai Shin Digital Infrastructure, CDIB Capital, and Shinetsu Science Industry, are building sovereign AI capacity in a country that has both the engineering talent and the industrial power infrastructure to support it.

This move sits inside a growing pattern. Andromeda AI hit a $1.5 billion valuation this week for on-demand GPU access, and the trend lifecycle data shows ”Sovereign AI Infrastructure” as a growing trend with momentum. The logic is structural: as AI models become critical business infrastructure, governments and enterprises are realizing that depending on a handful of U.S.-based cloud providers for AI compute is a geopolitical risk, not just a business one. The $12 billion number is not about building a data center. It is about building national AI independence.

One gigawatt is significant. For context, that is roughly the output of a nuclear power plant, dedicated to AI workloads. This kind of commitment makes sense only if the builders believe AI compute demand will remain intense for a decade or more. They are not betting on the next model release. They are betting on the permanent integration of AI into the global economy.

Here's what works: If your organization has significant AI compute needs, start mapping your supply chain for sovereignty risk. Where does your AI compute physically run? Which jurisdiction governs that data? If regulatory or geopolitical disruption cut off your current AI compute provider, what is your fallback? The companies building sovereign AI infrastructure are answering these questions now. If your organization depends on AI but has not asked them, you are exposed.

3. ”Buy High, Sell Higher Only Works in a Bubble.” A VC Legend Just Said It Out Loud.

Howard Morgan is not a commentator. He co-founded First Round Capital, the firm that seeded Uber, Square, and Notion before anyone knew what they were. When he tells the Economic Times that AI startup valuations are ”overheated”, the quote carries the weight of someone who has seen multiple cycles from the inside.

His argument is precise. The ”buy high, sell higher” playbook that has defined AI investing since 2023 only works if the next buyer is willing to pay more. When valuations detach from revenue, the chain breaks. Morgan is not saying AI is fake. He is saying the gap between what AI companies are valued at and what they earn is unsustainable. That gap has to close, and it closes in one of two ways: revenue catches up (good), or valuations come down (painful).

The data backs him up. ClickHouse just attracted talent at a $15 billion valuation, and the subtext of that hiring story is telling: the engineer who wrote about joining specifically cited ClickHouse's revenue trajectory, not its technology. In a market where even insiders are evaluating companies by their ability to generate cash rather than their ability to generate hype, the correction Morgan describes is already happening at the hiring level. Engineers are becoming the first price-correctors, choosing companies with real economics over companies with impressive demos.

Here's what works: If you are evaluating AI vendors, partnerships, or investments, apply Howard Morgan's filter: what is the revenue, not the valuation? A company valued at $15 billion with strong recurring revenue is fundamentally different from a company valued at $15 billion with a compelling pitch deck. Ask vendors for their net revenue retention rate. Ask partners about their unit economics. The companies that can show you numbers rather than narratives are the ones that will survive the correction.

4. The White House Is Trying to Govern AI by Friday. Yes, This Friday.

Axios reported that the White House is eyeing a Friday rollout for its AI framework, and the speed alone is a signal. This is not the result of years of deliberation. This is a government recognizing that the technology is outpacing every regulatory timeline anyone planned for. The framework aims to establish guardrails for federal AI use and set expectations for the private sector, all in a landscape where the rules are being written after the game has already started.

The same week, the House Financial Services Committee examined the Data Privacy Framework, adding another front to what is becoming a multi-pronged regulatory push. Privacy, AI governance, and data protection are converging in Washington at a pace that would have been unthinkable twelve months ago. Across the Atlantic, GDPR enforcement against generative AI in Europe has produced a track record that one analysis described as ”a lot of noise, one fine, zero survivors.”

The contrast between the U.S. and European approaches is instructive. Europe has comprehensive regulation with uncertain enforcement. The U.S. is building frameworks on the fly with uncertain scope. Neither has figured out how to regulate technology that changes faster than policy can be written. For enterprises operating in both markets, the practical implication is clear: you need to build to the strictest possible standard, because the floor keeps rising.

Here's what works: Do not wait for the White House framework to be finalized before acting. Map your organization's AI use cases against both U.S. and European regulatory expectations today. Build a compliance matrix that covers the strictest requirements across jurisdictions. The companies that treat compliance as a design constraint, built into the product from day one, will spend less than the companies that treat it as a retrofit project after the regulations land.

5. Wall Street's Smartest Quant Just Left Finance for an AI Lab. Here Is Why That Matters.

Jas Sekhon, Bridgewater Associates' chief scientist, is leaving the world's largest hedge fund to join Google DeepMind. This is not a retirement. It is a statement about where the most interesting problems in quantitative reasoning now live. When the chief scientist of a firm that manages $150 billion in assets decides that an AI research lab is a more compelling intellectual challenge, it tells you something about the center of gravity in applied mathematics.

The move comes at a moment when Demis Hassabis is expanding DeepMind's global reach, visiting South Korea to share AI development strategies with government officials. DeepMind is recruiting not just engineers but domain experts from the most quantitatively demanding fields in the world. Sekhon brings expertise in causal inference and econometrics, precisely the skills needed to make AI models that do not just correlate but actually understand cause and effect. The trend lifecycle data shows ”AI Talent Movement” as a growing trend, and this is what it looks like at the top: the best quantitative minds migrating from finance to AI because the problems are harder and the impact is larger.

This is the kind of talent movement that does not show up in headlines but reshapes industries. When finance loses its best scientists to AI labs, it means the AI labs are working on problems that finance cannot solve internally. It also means the AI models coming out of those labs in two to three years will be built by people who understand financial risk, causal reasoning, and complex system dynamics at a level that current AI models do not.

Here's what works: Track where the top scientific talent is moving, not just which companies are raising money. If your organization competes for quantitative talent with AI labs, acknowledge that you are in a different market than five years ago. The value proposition for elite researchers has shifted from ”solve hard problems and earn well” to ”solve the hardest problems in human history.” If you cannot offer that, offer what labs cannot: direct impact on real business decisions with immediate feedback loops.

6. This Startup Raised $14 Million to Replace AI Researchers with AI

Autoscience raised $14 million from General Catalyst, Toyota Ventures, and the Perplexity Fund to build what they describe as the world's first automated AI model building platform. Let that settle for a moment. AI that builds AI models. The recursion is no longer theoretical. It showed up in a seed round term sheet.

The investor list matters. General Catalyst does not write checks for science experiments. Toyota Ventures does not fund research that cannot be productized. And the Perplexity Fund signals that the search and reasoning infrastructure layer sees this as complementary, not competitive. Autoscience appeared in three separate articles this week, all independently covering the same thesis: the bottleneck in AI is not compute or data. It is the researchers who design, train, and iterate on models. And there are not enough of them.

This is the same recursive pattern we flagged when Siemens launched autonomous chip design agents. AI is becoming a tool in its own development pipeline. The implications are structural: if AI can meaningfully accelerate model development, the pace of AI improvement itself accelerates. That is not science fiction. That is a California startup with $14 million and a team that built their careers doing exactly this kind of work manually.

Here's what works: If your organization depends on ML engineering talent to build or fine-tune models, watch this space carefully. Automated model building will not replace your ML team next quarter. But within 18 months, the companies using these tools will iterate on models five to ten times faster than those relying entirely on human researchers. Start evaluating automated ML platforms now, not to replace your team, but to multiply their output.

7. Palantir Just Brought AI to Your Mortgage Application. Regulators Are Watching.

Palantir rolled out an AI-powered mortgage platform in partnership with a financial startup, and the move signals something that the rest of the financial services industry should pay attention to. This is not a chatbot answering customer questions. This is AI embedded in the decisioning layer of one of the most regulated consumer financial products in America.

The mortgage industry is a fascinating test case for enterprise AI because the regulatory surface is enormous. Fair lending laws, anti-discrimination requirements, truth-in-lending disclosures, and state-by-state compliance variations create a labyrinth that has historically made the industry resistant to technological change. Palantir is betting that its platform can navigate that labyrinth faster and more consistently than human underwriters, while simultaneously producing an audit trail that satisfies regulators. CIOs across industries are looking at AI to overcome M&A and integration complexity, and financial services is where the stakes for getting it right, or wrong, are highest.

The regulatory dimension makes this a leading indicator. If Palantir can deploy AI in mortgage origination without triggering regulatory pushback, it creates a template for every other regulated financial product: insurance underwriting, credit decisioning, investment advisory. If it stumbles, it becomes a cautionary tale that slows AI adoption in financial services by years. Either way, this is the moment when AI moves from the back office (analyzing data, generating reports) to the front office (making decisions that directly affect consumers).

Here's what works: If your organization operates in regulated financial services, track Palantir's mortgage deployment closely. Not because you need their specific product, but because the regulatory response will set precedent for every AI-powered financial decision your organization wants to make. Start documenting how your AI models make decisions now, before regulators require it. The companies that can show a clear decisioning audit trail will have a structural advantage when regulatory scrutiny intensifies.

Signal vs. Noise

🟢 Signal: Data Integration surged +1,988% in real influence across the ecosystem. Data Synchronization followed at +7,700%, and Data Quality rose +1,786%. Three foundational data disciplines, all exploding in structural importance at the same time. This is not a trend. This is the market correcting two years of skipping the boring stuff. Organizations that rushed to deploy AI models on messy, disconnected data are now doing the work they should have done first. The infrastructure layer is finally getting the investment the application layer got in 2024.

🟢 Signal: Workflow Automation rose +3,883% in structural influence. When automation starts gaining this kind of weight, it means enterprises are moving from ”let's experiment with AI” to ”let's make AI do actual work.” The shift from pilot projects to production workflows is the signal that separates real adoption from conference-talk adoption.

🔴 Noise: ”Agentic AI” appeared in both emerging AND declining trends simultaneously. Translation: the term is splitting. Serious practitioners are building real agent architectures (emerging). Marketing teams are slapping ”agentic” on existing products and calling it innovation (declining). When a buzzword appears on both sides of the lifecycle chart, it has entered the ”means everything, means nothing” phase. Judge companies by what their agents actually do, not by whether they use the word.

🔴 Noise: ”Regulatory Response to Generative AI” is declining despite 107 GDPR mentions in a single day. People are still writing about regulation. But the structural influence is fading because the conversation has shifted from ”how should we respond?” to ”we already know, now we need to implement.” The transition from regulatory theory to regulatory practice is happening faster than the thought leadership can keep up.

From the 190K

The Legal Infrastructure Nobody Is Building Headlines Around

We scanned 190,000 articles this week. Here is what no one is putting together:

”Copyright and Artificial Intelligence” was the single highest-rising foundational concept in our data. Four articles. Zero mainstream headlines. But the related concepts tell the real story: Collective Licensing, Opt-Out Exceptions, Digital Replicas, and Fair Use all appeared for the first time in the same period. Five legal concepts, all surfacing simultaneously, all connected to the same question: who owns AI-generated output?

This matters because copyright is the plumbing of the creative economy. Every AI company training on text, images, code, and music is operating in a legal grey zone that is about to get bright lines drawn through it. The European approach (opt-out exceptions, collective licensing schemes) and the American approach (fair use defenses, case-by-case litigation) are diverging, and the divergence will determine which types of AI training are legal, which require licensing fees, and which become prohibitively expensive.

The pattern is identical to what happened with data privacy ten years ago. In 2014, nobody was headlining GDPR. By 2018, it had reshaped every technology company's product roadmap. Copyright and AI is on the same trajectory, just earlier. The companies paying attention now are building licensing agreements and opt-out compliance into their training pipelines. The companies ignoring it are accumulating legal liability that will surface as litigation or regulation, probably both.

Below the surface: Data Lineage appeared across the ecosystem with one of the highest foundational importance scores in our data. Zero headlines. Here is how you spot real infrastructure: when something shows up everywhere but headlines nowhere, it means engineers are building on it and marketing has not caught up. Data Lineage is the audit trail that proves your data is trustworthy. Every compliance requirement, every AI governance framework, every regulatory disclosure depends on knowing where your data came from and what happened to it. Nobody writes headlines about lineage. But try passing an audit without it.

By The Numbers

• $120M for Oasis Security — Series B to build agentic access management. Sequoia, Cyberstarts, and Craft Ventures are betting AI agents need their own security stack.
• $12B GMI Cloud initiative — One-gigawatt sovereign AI data center complex in Japan. That is roughly the output of a nuclear power plant, dedicated to AI.
• $1.5B Andromeda AI valuation — On-demand GPU startup hits unicorn-plus status. Paradigm led the round.
• $14M for Autoscience — General Catalyst, Toyota Ventures, and Perplexity Fund backing AI that builds AI models.
• +7,700% Data Synchronization influence — The largest single-day influence surge in any infrastructure discipline this period. The plumbing is getting real.
• 107 GDPR mentions — Led all compliance frameworks, followed by CCPA (79) and HIPAA (58). The regulatory surface area keeps widening.
• $6M for PADO AI — AI-powered workload orchestration to maximize compute per megawatt. Solving the efficiency side of AI infrastructure.
• $15M for Rivia — AI engine for clinical trial data. Healthcare AI moves from diagnosis to drug development infrastructure.

Deep Dive: The Infrastructure Paradox (Why $12 Billion Bets and Bubble Warnings Can Both Be Right)

You know that moment when a club owner spends a fortune on a sound system? The DJs who play there might come and go. Some will pack the house. Others will clear the floor. But the sound system stays. It gets rented out. It becomes the foundation every act depends on. The club owner is not betting on any single DJ. He is betting on live music itself. That is exactly what is happening in AI right now, and it is the key to understanding why $12 billion infrastructure bets and ”valuations are overheated” warnings are not contradictory. They are describing different layers of the same market.

The Layer Cake of AI Investment

Howard Morgan's bubble warning and GMI Cloud's $12 billion commitment exist in different layers. Morgan is talking about application-layer companies: startups building AI products with impressive demos but unproven revenue models. Those valuations are detached from reality, and he is right. GMI Cloud is talking about infrastructure: power, compute, physical space. That investment will retain value regardless of which AI models or companies win, because every winner needs the same physical infrastructure. The confusion happens because the market prices them as if they are the same thing. They are not.

The Historical Rhyme

The dot-com bust did not destroy the internet. It destroyed the companies that could not generate revenue. But the fiber-optic cables, the data centers, the networking infrastructure that was built during the bubble? That became the foundation for everything that followed: Amazon Web Services, Netflix streaming, the entire cloud computing era. The companies that went bankrupt during the bust had, in many cases, built valuable infrastructure that survived them. The infrastructure outlasted the companies. It always does.

Who Survives the Correction

The filter is simple. Companies selling AI infrastructure (compute, power, security, data management) have customers who need their products regardless of which AI model is fashionable this quarter. Companies selling AI applications need to prove that their specific approach generates enough value to justify their specific valuation. The $120 million Oasis Security raised is infrastructure money: AI agents will need security regardless of whether the agents are built by a company valued at $10 billion or one valued at $100 million. The $14 million Autoscience raised is closer to the application layer, but the recursive nature of their product (AI that builds AI) puts them in a category where demand increases as the market grows.

What Actually Works

1. Separate your AI bets by layer: Infrastructure investments (data centers, security, data quality tools) have different risk profiles than application investments (AI-powered products, model providers). Evaluate them differently.
2. Ask the survival question: If AI valuations correct by 50% tomorrow, which of your AI vendors would still be in business? Infrastructure companies with utility-like revenue models survive corrections. Application companies burning cash to acquire customers might not.
3. Watch the talent flow: When Bridgewater's chief scientist leaves for DeepMind, it tells you where the next breakthroughs will come from. When engineers choose ClickHouse at $15B over AI startups at higher valuations, it tells you where the revenue is.
4. Build on infrastructure, not on hype: Every dollar spent on data integration, data quality, and security retains its value through a market correction. Not every dollar spent on the AI application of the month will.

The club owner who builds the best sound system does not worry about which DJ is hot this season. The system works for all of them. Build the system. Let others chase the DJs.

What's Coming

The White House AI Framework Will Force Enterprise Compliance Decisions Within 90 Days

The White House AI framework rollout scheduled for Friday will land in a market that has been waiting for regulatory clarity. Expect a scramble as enterprises evaluate whether their AI deployments comply with whatever guardrails the framework establishes. The companies that have already built compliance into their AI development process will have a head start. The companies that treated compliance as a future problem will discover it is a present one.

Copyright Litigation Will Reshape AI Training Economics

The simultaneous emergence of Collective Licensing, Opt-Out Exceptions, Digital Replicas, and Fair Use as rising concepts signals that the legal frameworks for AI training are being built right now. Within six months, expect landmark cases or legislative actions that establish whether AI companies must license training data. The impact will flow directly to model economics: if licensing is required, training costs increase, and the companies with proprietary data moats gain structural advantage.

AI Security Will Become a Board-Level Conversation

Oasis Security's $120 million raise, combined with Menlo Ventures' analysis of offensive AI reaching a tipping point, signals that AI security is transitioning from a technical concern to a governance requirement. Expect board directors and audit committees to start asking questions about AI agent access controls, policy chaining, and agentic risk management within the next two quarters. The companies that can answer those questions will have an easier time deploying AI at scale. The ones that cannot will face internal resistance from risk and compliance teams.

For Your Team

Monday's meeting prompt: ”Howard Morgan says AI valuations are overheated. GMI Cloud just committed $12 billion to AI infrastructure. Oasis Security raised $120 million for AI agent security. If these three data points are all correct simultaneously, what does that tell us about where our AI budget should actually go? Are we investing in infrastructure that retains value through a correction, or in applications that depend on the bubble continuing?”

The Infrastructure Layer Test:

1. Audit every AI investment by layer — Separate infrastructure spend (data quality, security, compute, integration) from application spend (AI-powered tools, model providers, chatbot implementations). Count the dollars in each bucket.
2. Apply the correction filter — For each application-layer investment, ask: if this vendor's valuation dropped 50% and they had to restructure, would your workflow survive? If no, you have concentration risk.
3. Score infrastructure completeness — Rate your data quality, data integration, security posture, and compliance readiness on a 1-5 scale. Any score below 3 is a gap that will cost more to fix later than to fix now.
4. Map your sovereignty exposure — Where does your AI compute physically run? Which jurisdictions govern that data? What happens if access to that compute is disrupted? If you cannot answer in one meeting, schedule a second one.

Share-worthy stat: Data Synchronization surged 7,700% in structural influence this period. Not AI models. Not chatbots. Not agents. The ability to keep data consistent across systems. The most invisible, most essential, most underloved capability in the entire stack is suddenly the fastest growing. The market is telling you something.

Go deeper: Track AI infrastructure signals and valuation trends in real-time

The Track of the Day

”Howard Morgan says bubble. GMI Cloud says $12 billion. Oasis Security says AI agents need their own security layer. Autoscience says AI can build AI. Bridgewater's chief scientist says the interesting problems are no longer in finance. And the single loudest signal in 190,000 articles? Data Synchronization surged 7,700%. Not the models. Not the agents. The plumbing. The market is telling you where the value is moving. The infrastructure always outlasts the hype cycle.”
Ins7ghts Knowledge Graph Analysis, March 2026

Today's set: ”Building a Mystery” by Sarah McLachlan. There is a line in that song about gathering the pieces and building something no one else can see. That is what the best data teams do in a market like this. While everyone argues about whether AI is a bubble or a revolution, the builders are quietly laying infrastructure, data integration, security layers, compliance frameworks, that will matter regardless of which side of the argument wins. The mystery is not whether AI is real. It is which layer of the stack captures the value. My money is on the plumbing. It always is.

Your DJ signing off. Separate your infrastructure bets from your application bets, secure your AI agents, and remember: the sound system outlasts the DJ. Every time.

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: March 20, 2026 | Curated by Yves Mulkers @ Ins7ghts