Pharma Giants Novo Nordisk & Eli Lilly Bet $2.6B on AI Drug Discovery

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. And the track that made me stop mid-mix was not another model release or another funding headline. It was a direction: Novo Nordisk partnered with an AI lab to transform how medicines are discovered and delivered, weeks after its rival Eli Lilly signed a similar deal. Meanwhile, Broadcom announced a multi-year partnership to build next-generation 2nm AI compute accelerators alongside a multi-gigawatt data center plan. Stanford released its AI Index 2026, revealing that more than half the world now uses AI but public trust is falling. And three startups you have never heard of raised $259 million to build infrastructure nobody headlines: optical switching, quantum acceleration, and engineering automation.

The Bottom Line: The biggest pharmaceutical companies are restructuring R&D around AI. Custom silicon is replacing generic chips. Public trust is dropping as adoption accelerates. And the real infrastructure bets are happening in companies most enterprise buyers have never heard of. The gap between what is happening and what is being reported is the story of April 2026.

The Tracks That Matter

1. Novo Nordisk and Eli Lilly Both Signed AI Drug Discovery Deals in the Same Quarter. The Pharma Arms Race Just Became Real.

Novo Nordisk signed a partnership to speed drug development using AI, marking one of the largest AI-pharma collaborations in the industry's history. The deal covers drug discovery, clinical trial optimization, and manufacturing process improvement. Coverage from multiple sources confirmed the scope: this is not a research experiment. This is a structural integration of AI into the drug development pipeline of the company that makes Wegovy and Ozempic.

What makes this significant is the timing. The race for AI-developed drugs is heating up, with rival Eli Lilly having sealed a similar deal weeks earlier. When the world's two largest pharmaceutical companies both sign AI partnerships in the same quarter, it stops being a trend and becomes an industry pivot. The estimated cost of bringing a drug to market is $2.6 billion and rising. AI promises to compress timelines from years to months. The companies that crack this first gain a structural advantage that compounds with every drug candidate.

Think of it like this: for decades, pharma R&D was the equivalent of a DJ spending hours in a record store, pulling one vinyl at a time, hoping to find the track that works. AI is the algorithm that scans the entire catalog and tells you which ten records the crowd will respond to before you even get to the venue. It does not replace the DJ's judgment. It changes what ”informed judgment” means.

Here's what works: If you are in healthcare, life sciences, or adjacent data services, map your organization's readiness for AI-assisted drug discovery partnerships. The pharma companies signing these deals will require data infrastructure from their suppliers, partners, and regulators. Ask your team: if a pharma partner required validated AI-ready datasets from us next quarter, could we deliver? If the answer involves the words ”we'd need to clean up our data first,” you have six months before that becomes a competitive disadvantage.

2. Broadcom Extends a Multi-Year AI Chip Partnership with Multi-Gigawatt Data Center Plans. Custom Silicon Just Became the Default.

Broadcom announced an extended partnership to deploy next-generation AI compute accelerators, a multi-year commitment that includes 2nm process technology. This is not a chip order. This is a structural bet on custom silicon as the future of AI infrastructure. The deal includes plans for multi-gigawatt data center expansion, which tells you everything about the scale of AI compute that hyperscalers now consider table stakes.

Multiple sources confirmed that Broadcom shares gained on the announcement, and the market reaction reflects a deeper shift. The era of buying off-the-shelf GPUs for AI workloads is ending for the largest players. Custom silicon, designed for specific AI architectures, delivers better performance per watt and per dollar at scale. The multi-year nature of the partnership signals that this is not a trial. It is a committed infrastructure strategy measured in gigawatts, not megawatts.

Last week we covered Credo's acquisition of DustPhotonics to own the silicon photonics layer. This week, Broadcom locks in custom silicon at 2nm. The pattern is unmistakable: AI infrastructure is disaggregating. Compute, connectivity, and cooling are becoming separate competitive layers. The companies that integrate all three will have an advantage that commodity hardware cannot match.

Here's what works: If you are evaluating AI infrastructure strategy, stop asking ”which GPUs should we buy” and start asking ”what does our total infrastructure stack look like in three years.” Custom silicon, optical interconnects, and power infrastructure are converging into a single strategic question. Ask your CTO: at our current growth rate, what is our data center power requirement in 2028? If the answer surprises your CFO, you need this conversation sooner.

---

Try It Yourself

---

3. Stanford's AI Index 2026 Shows Rapid Progress, Growing Safety Concerns, and Declining Public Trust. The Builders and the Users Are Splitting Apart.

Stanford's 2026 AI Index report revealed a striking paradox: AI capabilities are advancing faster than at any point in history, and public trust is declining at the same time. More than half the global population now uses AI tools regularly. Yet the people who build AI and the people who use it are developing fundamentally different views about its trajectory.

The report shows rapid progress alongside growing safety concerns. AI benchmarks are falling faster than researchers can create new ones. Model capabilities that were aspirational two years ago are now commodity features. But the same report documents a widening gap between AI insiders (who remain broadly optimistic) and the general public (who are increasingly skeptical). Stanford warns that AI insiders are losing the crowd, a framing that should alarm any enterprise leader planning AI deployments that require employee adoption.

This matters for a practical reason that has nothing to do with philosophy. Enterprise AI deployments depend on employee adoption. Employee adoption depends on trust. The report also shows China closing the gap with the US on AI capabilities, adding geopolitical pressure to a domestic trust problem. If your employees do not trust the AI tools you deploy, your ROI projections are fiction. The adoption curve bends on trust, not on capability.

Here's what works: Treat the Stanford AI Index like a market research report for your AI deployment strategy. The finding that matters most for your team: the trust gap between builders and users is widening. Before your next AI rollout, survey your employees on AI trust, not satisfaction. Trust predicts adoption. Satisfaction predicts nothing if people are not using the tool. And if your change management plan does not include a trust-building component, add one before Q3.

4. Supply Chain Attacks Are Moving Through Your SaaS Stack. And Researchers Just Mapped How to Attack AI Agents Through the Internet.

Vectra published an analysis of the rise of supply chain-driven data theft in SaaS environments. The attack pattern is deceptively simple: instead of targeting your systems directly, attackers compromise your vendor's vendor. The data they steal travels through authorized integrations, making it nearly invisible to traditional security monitoring. When your SaaS tools have API access to each other, your attack surface is not your perimeter. It is every vendor in your integration chain.

The same week, researchers mapped internet-based attacks on AI agents, cataloging how autonomous AI systems can be compromised through the web connections they use to gather information and take actions. This is the security story that most enterprise AI strategies are ignoring: every agentic AI system that browses the web, calls an API, or reads an email is a new attack surface. And these agents often have broader access permissions than the humans they replace.

A legal analysis of a recent AI code generation incident adds a third dimension: 512,000 lines of code generated in one night, zero permission requests. The legal frameworks for AI-generated code, AI agent actions, and AI-mediated data access are lagging years behind the technology. Thomson Reuters confirms that AI is scaling faster than the justice system can adapt, revealing a governance gap that widens with every autonomous system deployed.

Here's what works: Conduct a supply chain audit of your SaaS integrations this month. For every tool in your stack, answer three questions: what data does it send to other services via API? What data does it receive? Who authorized each connection and when was it last reviewed? If your CISO cannot answer all three for your top 10 SaaS tools, your supply chain security is based on trust, not verification.

5. The EU AI Act Has a Runtime Problem. Most Teams Checked the Compliance Boxes at Deployment and Missed the Ongoing Obligations.

An analysis from Opaque Systems revealed that most teams are not aware of the EU AI Act's runtime compliance requirements. The distinction matters: compliance at deployment time (static checks, documentation, risk assessments) is where most organizations focused their EU AI Act preparation. But the regulation also requires ongoing monitoring, real-time logging, and continuous risk reassessment for high-risk AI systems. Most teams do not know they have this obligation because it looks different from every other compliance requirement they have managed before.

The EU Digital Omnibus directive adds another layer, consolidating cybersecurity, data protection, and AI governance into a single regulatory framework that demands operational compliance, not just documentation. A parallel analysis of recent compliance fines in the US shows that regulators on both sides of the Atlantic are shifting from warning to enforcement. The pattern across fines: they target organizations that had policies on paper but not in practice.

Our monitoring tracked 108 GDPR references, 69 CCPA references, and 62 HIPAA mentions in a single day across the article corpus. When regulatory language density stays this high across multiple frameworks simultaneously, it signals that enforcement infrastructure is being built, not just discussed. A comprehensive guide to AI governance frameworks cataloged more than a dozen regulatory frameworks now active globally. The organizations that built compliance into their architecture are absorbing this. The ones that bolted it on afterward are about to learn the difference.

Here's what works: Ask your legal team one question: do we have runtime monitoring for our high-risk AI systems, or did we stop at deployment-time compliance? If the answer is deployment-time only, you are compliant with the EU AI Act as it was understood in 2025, not as it is enforced in 2026. Budget for continuous AI monitoring tooling before your next audit cycle. The cost of retrofitting runtime compliance is an order of magnitude higher than building it in.

6. Three Startups Raised $259 Million This Week to Build AI Infrastructure Nobody Headlines. One Uses Light. One Uses Quantum. One Uses Agents.

nEye.ai raised $80 million to scale AI optical switching. If that sentence did not make you sit up, here is why it should: optical switching determines how fast data moves between AI processors. As AI clusters scale from hundreds to thousands of chips, the interconnect layer becomes the bottleneck. nEye.ai is building the traffic management system for AI data centers. Last week, Credo acquired DustPhotonics for the same reason. Two optical infrastructure deals in two weeks is a signal, not a coincidence.

Sygaldry Technologies raised $139 million to develop quantum-accelerated AI infrastructure. The thesis: classical computing cannot keep scaling AI workloads indefinitely, and the companies that build the bridge between quantum hardware and AI software will capture the next generation of infrastructure spending. Meanwhile, German startup Synera landed $40 million to automate engineering workflows with AI agents, bringing agentic AI to mechanical and industrial engineering design.

None of these three companies will trend on LinkedIn this week. But $259 million in infrastructure funding in a single week, all targeting layers beneath the AI models themselves, tells a story the headlines are missing. The AI industry is disaggregating. The model layer got the first wave of capital. The infrastructure layer is getting the second. The companies that connect compute, move data at the speed of light, and bridge classical and quantum processing will capture margin that model providers cannot.

Here's what works: Add infrastructure-layer startups to your AI vendor watch list. If your current evaluation focuses only on model providers, you are looking at one layer of a five-layer stack. Ask your infrastructure team: who are the three most important companies in our AI stack that we do not have a direct contract with? Those are your hidden dependencies, and someone just raised $80 million to compete with or replace them.

Signal vs. Noise

🟢 Signal: Pharma AI partnerships are becoming structural investments, not pilot programs. Novo Nordisk signed a partnership to transform drug discovery and delivery weeks after Eli Lilly sealed a similar deal. When the race for AI-developed drugs draws coverage from seven sources in a single day, this is not a proof of concept. This is an industry restructuring R&D around AI at the highest levels of commitment and capital.

🟢 Signal: AI infrastructure capital is shifting from compute to connectivity. nEye.ai raised $80 million for optical switching one week after Credo acquired DustPhotonics for silicon photonics. Two optical infrastructure deals in two consecutive weeks means the smart money has moved past the GPU shortage narrative. The next bottleneck is how data moves between processors, and the companies solving it are raising serious capital.

🔴 Noise: ”Declining public trust in AI” as a reason to slow down. Stanford's AI Index shows trust is falling, but adoption is accelerating. More than half the global population uses AI. The trust gap is real, but it does not predict slower adoption. It predicts messier adoption. The organizations that treat the trust gap as a deployment problem (change management, transparency, employee communication) will fare better than those that treat it as a reason to wait.

From the 190K

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

Three separate companies launched ”agentic” capabilities on the same day, across three different product categories.

Qlik brought agentic execution to data engineering. Teradata launched an Analyst Agent on a major cloud marketplace. And Diginomica covered Qlik's approach with a headline that says everything: ”Qlik's most important AI feature is knowing when to say nothing.” Three product launches, three categories (data engineering, analytics, enterprise platforms), one shared vocabulary.

When a label migrates from chatbot marketing to data pipeline engineering in a single product cycle, it stops being a buzzword and starts being a product category. ”Agentic AI” appeared in chatbot contexts for most of 2025. This week, it showed up in data integration, query optimization, and workflow automation. The label is the same. The application layer has shifted entirely.

🔍 Below the surface: A German startup raised $40 million to bring AI agents to mechanical engineering workflows. Synera is not building chatbots. It is building agents that automate engineering design processes. When the ”agentic” label reaches industrial engineering, it has moved from software convenience to operational infrastructure. Watch for ”agentic” to appear in supply chain, manufacturing, and logistics contexts within 90 days.

By The Numbers

• $259 million: Combined funding this week for three AI infrastructure startups (nEye.ai, Sygaldry, Synera) building beneath the model layer. Optical switching, quantum acceleration, and engineering automation.
• 7 sources: The number of publications that covered Novo Nordisk's AI drug discovery partnership on a single day. When a pharma deal generates more coverage than a model release, the industry center of gravity is shifting.
• 2nm: The process node for Broadcom's next-generation custom AI accelerators. Multi-year commitment, multi-gigawatt data center scale. Custom silicon is replacing commodity chips for the largest AI workloads.
• 108 GDPR references: Compliance mentions in a single day across our monitoring corpus. CCPA hit 69, HIPAA hit 62. Regulatory language density remains elevated across every major framework simultaneously.
• $680,000: What one pharmaceutical company spent building a ”comprehensive data governance strategy” before abandoning it after 18 months. Their competitor solved a single compliance gap and achieved working governance in 90 days.
• 50%+: Share of the global population now using AI regularly, per Stanford's AI Index 2026. Adoption is accelerating. Trust is declining. Both facts are true at the same time.
• 20% decline: Drop in employment for young software developers since 2022, according to the Stanford AI Index. The job market is absorbing AI capabilities faster than workforce planning can adapt.

Deep Dive: The Infrastructure Inversion, or Why the Bottleneck Keeps Moving and Most Strategies Are Solving Last Quarter's Problem

You know that moment at a festival when the bass drops and half the crowd pushes forward, only to realize the real energy is coming from a stage they did not know existed behind them? That is what is happening in AI infrastructure right now. Everyone is watching the main stage (model releases, funding rounds, capability benchmarks) while the real action has moved to stages most people have not noticed yet.

The Bottleneck Moved, and It Moved Fast

Twelve months ago, the AI infrastructure conversation was dominated by GPU supply. Could you get enough chips? From which vendor? At what price? That conversation is still happening, but it is no longer the binding constraint for the companies at the frontier. Broadcom's multi-year chip partnership at 2nm tells you where the hyperscalers have moved: custom silicon, designed for specific workloads, at a scale measured in gigawatts. The GPU shortage narrative is last year's problem for the companies writing the biggest checks.

The New Constraints Are Invisible

The bottleneck has moved to three layers that most enterprise AI strategies do not address. First, connectivity: nEye.ai raised $80 million for optical switching because data movement between processors is now slower than the processors themselves. Second, governance: the EU AI Act's runtime problem means compliance is no longer a one-time checkpoint but a continuous obligation most teams are not staffed for. Third, trust: Stanford's data shows that the people using AI tools are losing confidence in them even as the tools improve. Each of these constraints operates below the level of visibility that most strategy documents address.

The Inversion Is the Opportunity

Here is the pattern that only emerges when you look at all three simultaneously. The companies that are winning the infrastructure race are not the ones solving the most visible problem. They are the ones solving the next problem. Broadcom moved to custom silicon before generic GPUs became abundant. nEye.ai moved to optical switching before bandwidth became the headline constraint. Sygaldry raised $139 million for quantum AI infrastructure before quantum computing entered most enterprise roadmaps. The infrastructure inversion rewards the companies that see where the bottleneck is going, not where it is.

What Actually Works

1. Audit your infrastructure stack, not just your model selection. The model is one layer. Compute, connectivity, governance, and trust are four more. If your AI strategy document only addresses model selection, it covers 20 percent of the decisions that will determine success.
2. Budget for runtime compliance, not just deployment compliance. The EU AI Act, GDPR, and emerging US state regulations all require ongoing monitoring. Build the cost of continuous compliance into your AI budget from day one.
3. Track infrastructure-layer startups alongside model providers. The companies raising capital for optical switching, quantum acceleration, and agentic automation today will be the vendors you depend on in 18 months. Start evaluating them now, before your procurement team has to scramble.
4. Measure trust, not just adoption. Stanford's data proves that high adoption and low trust can coexist. If your AI deployment metrics track usage but not employee confidence, you are measuring the accelerator without checking the brakes.

The DJ who reads the room knows that the energy of a set is not determined by the loudest track. It is determined by the transitions between tracks. The AI infrastructure landscape is in the middle of a transition most strategies have not accounted for. The bottleneck has moved. The question is whether your strategy moved with it.

What's Coming

Quantum AI Convergence Will Force Enterprise Roadmap Updates by Q4

Sygaldry Technologies raised $139 million for quantum-accelerated AI infrastructure the same week BMO Financial expanded its quantum and AI strategy through new partnerships. When both startups and major banks are investing in quantum AI simultaneously, the ”quantum is five years away” talking point has an expiration date. Expect enterprise quantum pilot programs in optimization and risk modeling to accelerate in H2 2026.

AI Drug Discovery Results Will Create a New Benchmark for AI ROI

With Novo Nordisk and Eli Lilly both committing to AI-powered drug discovery, the pharma industry is about to generate the first large-scale data on whether AI can measurably compress drug development timelines. Early results from these partnerships will set expectations for AI ROI across every industry. If pharma delivers a provable timeline compression, every enterprise AI business case will be rewritten to match.

Agentic Data Engineering Will Become a Standalone Product Category

Qlik, Teradata, and Synera all launched agentic capabilities this week across data engineering, analytics, and industrial design. Expect analyst coverage to formalize ”agentic data engineering” as a distinct category before Q3. The tools that move data, clean data, and transform data are about to do it autonomously.

For Your Team

Wednesday's meeting prompt: ”For every AI deployment in our pipeline: do we have runtime monitoring for ongoing compliance, or did we stop at the deployment-time checklist? If our AI systems are operating right now without continuous compliance monitoring, we are accumulating regulatory exposure with every hour they run.”

The Infrastructure Stack Audit:

1. Model layer. Which models do we use, who provides them, and what is their profitability? An unprofitable model provider is a risk position.
2. Compute layer. Are we on commodity hardware or custom silicon? At our scale, does custom silicon change the economics?
3. Connectivity layer. When was our data center interconnect last upgraded? Is optical switching on our roadmap?
4. Governance layer. Do we have runtime compliance monitoring or deployment-time documentation only? Who owns continuous AI compliance?
5. Trust layer. Have we measured employee trust in our AI tools, or only adoption rates? Trust predicts sustained use. Adoption without trust predicts eventual rejection.

Share-worthy stat: More than half the global population now uses AI regularly. Public trust in AI is declining. Both are true at the same time. The gap between adoption and trust is the defining risk of 2026.

Go deeper: Track AI infrastructure and governance signals in real-time →

The Track of the Day

”AI insiders are losing the crowd.”
Stanford HAI, AI Index 2026

Today's set: ”Losing My Religion” by R.E.M. In 1991, Michael Stipe sang about the corner of a view, the spotlight finding everyone but him. Thirty-five years later, AI insiders are having their own moment. They built something extraordinary. More than half the world adopted it. And somewhere along the way, they lost the trust of the people using it. The technology keeps improving. The faith in it does not. That gap, between what AI can do and whether people believe it should, is the track that will define this year. The DJ cannot play a set the crowd does not trust.

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: April 15, 2026 | Curated by Yves Mulkers @ Ins7ghts