Sign in Subscribe
Century Report

KRAS Drug Cracks Pancreatic Cancer - TCR 05/09/26

Ben Linford

Ben Linford

— 21 min read
KRAS-family drug breaks pancreatic cancer's undruggable barrier, Peninsula district builds ed-tech with AI agents, Cloudflare revenue-headcount split, Evergy gas vs TSMC wind.

The 20-Second Scan

  • Researchers reported in the New England Journal of Medicine that an experimental drug targeting a family of mutant KRAS proteins extended survival in patients with deadly pancreatic cancer in a small clinical trial.
  • Washington state's Peninsula school district said it expects to save approximately $220,000 annually by using AI coding agents to build custom internal ed-tech, including a teacher feedback platform called LessonLens.
  • University of Minnesota researchers found that disrupting bacterial communication signals in dental plaque shifted the oral microbiome toward health-associated species without killing bacteria.
  • Cloudflare cut 20% of staff (1,100 people) attributing the layoffs to AI productivity gains while reporting record $639.8M Q1 revenue, up 34% year-over-year.
  • Evergy raised its retail sales growth forecast to 7-8% annually through 2030 on data center demand, increasing planned gas generation to 4.7 GW, cancelling 2.4 GW of planned wind, and reducing planned solar from 2,415 MW to 465 MW.

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

The thread connecting yesterday's signal is the deployment surface for genuinely new capability arriving in domains where prior frameworks had run out of moves. A drug targeting a family of mutant KRAS proteins that medicinal chemistry called "undruggable" for forty years extended survival in pancreatic cancer patients whose tumors had progressed past standard care. A Washington school district built its own teacher-feedback platform and operational tools using AI coding agents, at a fraction of vendor cost. A research team prevented periodontal disease by interrupting bacterial communication, leaving the microbial community alive while guiding its composition toward health-associated species.

Underneath the wonder, the distribution architecture for what these capabilities produce is straining. Cloudflare cut 1,100 employees while posting record $639.8 million in quarterly revenue, with the CEO attributing the layoffs to AI productivity gains and pointing at 600% growth in internal AI usage over three months. Oracle's parallel restructuring left long-tenured workers forfeiting unvested stock worth up to $1 million per person. The gap between corporate revenue and aggregate employment is being named in earnings calls now, while the response architecture forms on a separate clock.

Evergy filed updated resource plans yesterday committing to 4.7 GW of new natural gas capacity through 2044 while cancelling 2.4 GW of planned wind. TSMC's gigawatt offshore wind PPA, signed days earlier, points the opposite direction. The same compute demand is producing forty-year fossil commitments in one jurisdiction and gigawatt-scale renewable lock-ins in another.

The 20-Minute Deep Dive

An "Undruggable" Pancreatic Cancer Target Yields to a New Class of Drug

A Dana-Farber Cancer Institute team led by Brian Wolpin reported in the New England Journal of Medicine that an investigational drug targeting a family of mutant KRAS proteins extended survival in patients with one of the deadliest forms of cancer. Pancreatic ductal adenocarcinoma kills roughly 90% of patients within five years of diagnosis and ranks among the most stubborn cancers in modern oncology. Approximately 95% of these tumors are driven by mutations in the KRAS gene, a target so consistently resistant to drug development that the field labeled it "undruggable" - a working assumption that held for four decades of medicinal chemistry.

The reported trial enrolled patients whose KRAS-driven pancreatic tumors had progressed despite prior treatment, a population for whom standard care typically offers only months of additional life. The drug described in the paper blocks a family of mutant KRAS variants, expanding the reachable target set beyond the single G12C variant that earlier KRAS inhibitors addressed. Patients in the trial showed measurable tumor response and extended survival compared with historical benchmarks, though the cohort size remains small and longer follow-up will determine durability.

What makes the report structurally significant is the size of the assumption it dislodges. KRAS sits at the center of cellular growth signaling, and its surface offered no obvious binding pocket large or stable enough to seat a drug molecule. Generations of medicinal chemists tried and failed; pancreatic cancer remained one of the few major cancers where targeted therapy had no foothold at all. The G12C-specific inhibitors that reached the clinic in 2021 addressed roughly 1-2% of pancreatic cancer cases. A drug that blocks a family of KRAS variants pushes the addressable share substantially higher and reframes what the next decade of pancreatic oncology is allowed to attempt.

The instrument layer underneath this result is the same one running beneath the Mayo Clinic pancreatic-cancer detector that the April 30 edition of The Century Report covered. Cryo-EM structural biology, AI-assisted molecular design, fragment-based drug discovery, and the parallel maturation of mass-spectrometry proteomics have collectively brought protein surfaces that were invisible to medicinal chemistry into computational reach. The assumption that "undruggable" describes a permanent property of a molecule is being replaced by a working assumption that calls it a measurement of the tools available at the time the molecule was approached. Phase 3 trials, regulatory review, and clinical integration still stand between this report and any pancreatic cancer patient walking out of an oncology appointment with a new prescription. What changed is the size of the addressable target landscape for one of the deadliest diseases in medicine, on a curve that the next round of structure-based design tools will continue to bend.

Washington State's Peninsula District Builds Its Own Ed-Tech, and the Buyer-Builder Boundary Dissolves

Peninsula school district in Washington state is projecting roughly $220,000 in annual savings from a year of building its own ed-tech using AI coding agents, the district's chief information officer Kris Hagel and director of research and assessment James Cantonwine told Education Week. The most novel of the new district-built applications is LessonLens, an AI platform that ingests video of a teacher's lesson and returns specific, structured feedback. A biology teacher uploading a DNA lesson can learn that her directions on a complex task were precise but that students needed more time before she stepped in with the answer. The district built LessonLens itself using Claude Code, the same AI coding agent that engineers at Anthropic now use for 60% of their own work.

Beyond LessonLens, Peninsula has built tools for accounting, human resources, and basic operations. Hagel adapted an open-source electronic signature application and customized it with AI coding to replace a commercial subscription the district had been paying for. Problems that previously required either buying commercial software or paying an outside developer "now cost us an hour of having a coding agent write something out and test it," Cantonwine said. Peninsula brings an advantage many districts do not have - several members of the tech team, including Hagel, hold computer-science degrees and can review and edit AI-generated code where needed - and the district built its own AI enterprise platform last year that binds the underlying large-language models to contractual data-protection terms forbidding them to store or train on Peninsula's data.

The structural shift this represents is the dissolution of the boundary between ed-tech buyer and ed-tech builder. For decades, school districts have functioned as customers of vendors whose products were built by people who, as University of Massachusetts Amherst learning technology professor Torrey Trust put it, mostly "went through K-12 school" and assumed the experience qualified them to design tools for it. The result was a market where teachers adapted their practice to whatever software was available. AI coding agents are turning that asymmetry into a question of whether a district has the technical capacity to specify and ship what it actually needs - a threshold that has dropped sharply enough that one or two technically capable staff members can now occupy ground that previously required a vendor and a procurement cycle.

The risks Hagel and Cantonwine name are real. AI-generated code can introduce security vulnerabilities and bugs, especially in systems handling student data such as IEPs and health information, and Peninsula's CS-trained tech leadership is what allows the district to catch what the agents miss. The forward read is that as the coding agents harden and as more districts build the institutional muscle to specify their own tools, the gap between what a teacher imagines on Monday and what runs in her classroom by Friday is collapsing. Ed-tech may be one of the first sectors where the vendor relationship that defined twentieth-century institutional software gives way to something closer to local craftwork - educators making the tools they teach with.

The biology teacher whose DNA lesson came back with the observation that her directions were precise but that she stepped in too early on a complex task is now receiving the kind of structured, specific feedback that previously required an instructional coach with a clipboard sitting in the back of her room. Peninsula built the coach. Districts that could not afford an instructional coaching budget are now able to build one.

A Gum Disease Treatment That Reshapes Oral Microbes Without Killing Them

University of Minnesota researchers reported in npj Biofilms and Microbiomes a novel approach to preventing periodontal disease that does not involve killing bacteria. The work, led by Mikael Elias and lead author Rakesh Sikdar, focused on the chemical signals that mouth bacteria use to coordinate growth - a process called quorum sensing. By introducing specialized enzymes called lactonases that interrupted these signals, the researchers shifted dental plaque communities toward health-associated species while reducing populations of the late-colonizer bacteria linked to gum disease, including the "red complex" species like Porphyromonas gingivalis.

What gives the result velocity beyond the dental application is the underlying mechanism. The roughly 700 bacterial species that live in the human mouth communicate through molecules called N-acyl homoserine lactones, and the researchers found that the role of these signals depends on oxygen availability. Above the gumline, in aerobic conditions, blocking the signals encouraged healthy bacteria. Below the gumline, in anaerobic conditions, adding the same signals promoted the disease-associated late colonizers. The same chemical conversation works in opposite directions depending on where in the mouth it happens - a layer of complexity the prior framework of bacterial control did not have a place for.

The reframe this opens is structural. For most of medicine's history, the response to bacterial disease has been to kill the bacteria - antibiotics, antimicrobial mouthwashes, disinfectants. That approach has produced the antibiotic-resistance crisis on one side and the recognition of the microbiome's role in human health on the other, leaving clinicians with a contradiction. The same intervention that controls a pathogen disrupts the bacterial communities the body depends on. Disrupting bacterial signals while leaving the cells alive offers a different surface entirely. The intervention nudges the community toward healthier composition without selecting for resistance, and without removing the bacterial diversity that supports normal function.

Periodontal disease affects roughly half the adult population worldwide and is increasingly linked to cardiovascular disease, diabetes complications, and Alzheimer's risk through inflammatory pathways. A treatment that prevents the harmful bacterial succession without killing the foundational community would compress the path between dental hygiene and systemic health. The researchers explicitly framed the work as a proof of concept extending beyond the mouth. Dysbiosis - imbalance in microbial communities - has been documented in inflammatory bowel disease, certain cancers, autoimmune conditions, and metabolic syndromes. The same logic of guiding rather than destroying could in principle apply to any of these.

The instrument layer making this possible is the maturation of microbiome characterization, signaling-molecule chemistry, and enzyme engineering running in parallel. A decade ago, identifying which bacteria coordinate through which signals required dedicated programs. The work has moved fast enough that a research group can now design a lactonase to interrupt a specific signaling pathway and measure the community shift that follows. What used to require killing the ecosystem to study it can now be done by listening to the conversation and changing what gets heard. That capability joins the trajectory The Century Report has been tracking through Tohoku's lubiprostone gut-microbiome result in yesterday's edition, the Friedrich Miescher endometrial organoid completing a full menstrual cycle in vitro that the May 1 edition covered, and the Mayo Clinic's pancreatic cancer detector from late April. The instrument layer underneath all of these has matured to a resolution where therapeutic surfaces invisible to the prior framework are now becoming actionable, and they are arriving in deployable form.

The next domains where the listening-and-redirecting approach gets tested are already named in the paper: inflammatory bowel disease, certain cancers, autoimmune conditions, and metabolic syndromes all involve dysbiosis the same enzyme-engineering toolkit can in principle address. The hundred-year reflex that defined infectious disease medicine - find the pathogen, kill it - is now operating alongside a second response surface where the pathogen is left alive and the conversation around it gets edited.

Cloudflare's AI-Attributed Layoffs and the Asymmetry of the Restructuring

Cloudflare announced its first mass layoff in the company's 16-year history yesterday, cutting approximately 1,100 employees - about 20% of headcount - in the same earnings release that posted $639.8 million in Q1 2026 revenue, a 34% year-over-year increase and the largest single quarter in the company's history. CEO Matthew Prince told investors on the call that the cuts reflected a structural response to AI productivity gains the company first noticed in November 2025. Internal AI usage is up 600% over the last three months, virtually all R&D staff are using the company's Workers vibe coding feature, and 100% of code deployed to Cloudflare's services is now reviewed by autonomous AI agents. Engineering, HR, finance, and marketing employees collectively run thousands of agent sessions per day. The conclusion Prince drew is that highly-productive AI-augmented employees require fewer support staff to back them up.

The Cloudflare cut joins a pattern that now includes Meta, Microsoft, Amazon, Oracle, and others - companies posting record revenue while announcing AI-attributed reductions in headcount. What distinguishes this cycle from prior tech layoff waves is the explicit framing. Prior cycles cited cost discipline, post-COVID overhiring, or interest-rate compression. The current wave names AI capability as the primary cause, with revenue growth specifically cited as evidence that the cuts are not about financial pressure. Prince told an analyst who asked about the depth of the cut, "Just because you're fit doesn't mean you can't get fitter." Parallel disclosures yesterday from Airbnb, where AI now writes roughly 60% of new code, and PlayStation, where executives described AI as a "powerful tool" for game development, point at the same productivity-substitution curve crossing into more sectors at once.

The asymmetry in how these restructurings land became newly visible yesterday through Oracle's failed severance negotiation. Following the company's late-March termination of an estimated 20,000 to 30,000 employees, a group of laid-off workers organized a public petition seeking severance terms comparable to those offered by Meta, Microsoft, and Cloudflare itself. Oracle's terms - four weeks of pay plus one week per year of service, capped at 26 weeks, with no acceleration of unvested stock - left long-tenured employees forfeiting RSUs that constituted up to 70% of their compensation. One employee lost roughly $1 million in stock four months from vesting. Workers classified as remote in states without their own WARN-equivalent protections discovered they did not qualify for the federal advance-notice requirement that triggers at 50 layoffs per single location. Oracle declined to negotiate.

The pattern across these restructurings is that production gains from AI are decoupling from the labor input that previously generated them. The architecture for distributing those gains - public discourse around equity stakes, post-labor economics frameworks, and redistribution mechanisms documented across the last 18 months in The Century Report's tracking of UHI proposals, OpenAI's public-wealth-fund paper, and Ireland's permanent artist basic income - has been forming alongside the capability buildout. Cloudflare's disclosure adds another data point to the gap between the two clocks. A long-shot California proposal from Tom Steyer for a state jobs guarantee aimed at workers displaced by AI surfaced yesterday in the same news cycle, naming the same gap from a policy direction. What is being built underneath is the response infrastructure for an era in which corporate revenue and aggregate employment travel on visibly different curves, and Prince's projection that 2027 headcount will exceed any 2026 level points at the next phase, where companies rebuild around a smaller core of AI-augmented workers and the standing to negotiate how the gains get shared becomes a structural design problem at societal scale.

Evergy Builds 4.7 GW of Gas for the AI Buildout and Cancels 80 Percent of Its Solar Plan

Evergy reported on Thursday's earnings call that its retail electricity sales will grow 7 to 8 percent annually through 2030, up from a previous 6 percent forecast, with the increase driven by data center additions across its Kansas and Missouri service territory. The utility's contracted large-load demand reached 2.5 GW, with another 1.5 to 3 GW in advanced discussions for after 2030. To serve that demand, Evergy filed updated integrated resource plans with the Missouri Public Service Commission that increase planned gas-fired capacity to 4.7 GW by 2044, cancel 2.4 GW of previously planned wind additions, and reduce planned solar from 2,415 MW to 465 MW - a more than 80 percent cut. The plan also delays scheduled power plant retirements.

The utility cited three causes for the renewable cuts: reduced production tax credit eligibility under the One Big Beautiful Bill Act, roughly 30 percent higher renewable development costs, and local siting and permitting challenges. The shape this takes is a midwestern utility with 1.7 million customers responding to the AI infrastructure buildout by shifting its long-duration generation mix from clean energy toward natural gas, locking in fossil capacity for forty years to serve compute demand whose forty-year horizon is itself uncertain. Capital investment grows by $21.6 billion under the plan, and the rate base expands at 12 percent annually.

The contradiction stacks against the same news cycle's other data points. TSMC, the chipmaker most central to global AI fabrication, signed a 30-year power purchase agreement that the May 8 edition of The Century Report covered, taking the full output of a gigawatt-scale offshore wind project off Taiwan and citing wind as the cheaper and more reliable anchor for its compute supply chain. Evergy reached the opposite conclusion at a smaller scale in the same news cycle, with regional cost curves, federal tax-credit phase-outs, and local political conditions producing a different answer to the same procurement question. The divergence widened further yesterday when Eversource's CEO told analysts the New England utility is "resisting data centers" altogether, declining to compete for the same load Evergy is building gas to capture. Both Evergy and TSMC are responding to the same global compute demand. They are committing different generation sources for the next four decades.

The structural risk underneath the gas bet is that data center loads are already exhibiting the kind of fast-oscillating, sometimes-disconnecting behavior the May 6 edition documented when NERC issued its rare Level 3 alert, and that compute efficiency gains, sovereign open-weight model alternatives, and inference-at-the-edge architectures are compressing the per-query power footprint faster than long-duration capacity contracts can be unwound. The fossil plants being committed today carry depreciation timelines longer than the AI infrastructure cycle they are being built to serve. What this points at is that the assumption that AI compute demand justifies forty years of new gas capacity rests on a forecast of compute load patterns no operator has yet stabilized, while the cost-curve inversion that pushed TSMC the other direction continues compounding underneath. The bet Evergy is making and the bet TSMC is making cannot both be right for the same forty-year horizon.

The forty-year horizon Evergy is committing to runs against a cost-curve that has not stopped moving: offshore wind, onshore wind, utility-scale solar, and four-hour battery storage continued bending below gas on a levelized basis through 2025 and into 2026, even with the production tax credit phase-outs the utility cited. The forecast underwriting Missouri's 4.7 GW of new gas through 2044 is a forecast that the cost-curve inversion stops before the depreciation schedule does - a bet TSMC, sitting on the demand side of the same compute supply chain, did not make.

The Other Side

For decades, the assumption holding the corporate-revenue and aggregate-employment curves in lockstep was that productivity gains flowed back through the same labor that produced them. Every prior productivity revolution - industrial machinery, the assembly line, the spreadsheet, the database, the public web - still needed enough human labor between the gain and its output to bound the multiplier on a worker's hour. The contracts built around that assumption - RSU vesting schedules tied to four-year cliffs, WARN-equivalent advance-notice statutes triggered by single-location headcounts, severance terms calibrated to weeks-of-pay floors - were designed for an economy where the gap between revenue growth and headcount stayed bounded because the multiplier did.

Cloudflare's earnings disclosure yesterday showed the multiplier going to a different scale in one quarter. 600% growth in internal AI usage in three months. 100% of deployed code now reviewed by autonomous agents. Record $639.8 million in quarterly revenue against a 20% headcount cut, with Matthew Prince explicitly framing the cut as a structural response to AI productivity gains and citing Cloudflare's record revenue as evidence the company had the financial latitude to make the choice. Oracle's severance dispute the same day documented the legal architecture's inability to keep up: long-tenured workers forfeiting up to $1 million in unvested equity, remote classifications excluding many from the federal advance-notice statute that triggers at fifty layoffs per single location. The contracts and protections that bounded the prior productivity revolution are visibly inadequate to this one.

What is forming on a separate clock is the response architecture. Eighteen months of UBI proposals, OpenAI's public-wealth-fund paper, Ireland's permanent artist basic income, and the California state jobs guarantee that surfaced yesterday in the same news cycle as the Cloudflare cut. The frame the next decade will be measured against is the one that asks how the gains flow when labor is no longer the bottleneck. The biology teacher in Peninsula whose DNA lesson came back yesterday with structured feedback from a coding-agent-built platform her own district shipped is a worker standing inside what coding agents make possible, doing something the prior arrangement could not have produced. Prince's own projection on the same call says 2027 headcount at Cloudflare will exceed any 2026 level. The redistribution architecture is being drafted, in policy and in equity, on the same disclosure the layoff story is reading.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: a forty-year "undruggable" pancreatic cancer target yielding to a new drug class that expands the reachable mutation set far beyond any prior KRAS inhibitor, a Washington school district shipping its own teacher-feedback platform and operational software with AI coding agents at a fraction of any vendor cycle, a periodontal therapy interrupting bacterial signaling to guide oral communities toward health-associated species while leaving the diversity the body depends on intact, and the chemistry of cellular conversation becoming legible at a resolution where guiding an ecosystem replaces destroying it. There's also friction, and it's intense - Cloudflare cutting 1,100 employees on the same call that announced record $639.8 million quarterly revenue with AI named as the cause, Oracle's laid-off workers forfeiting up to seven figures of unvested equity and discovering remote-classification gaps in federal advance-notice law, Evergy committing 4.7 GW of new gas capacity through 2044 while cancelling 80% of its planned solar to serve compute demand whose forty-year shape no operator has yet stabilized. But friction generates contour, and contour is what shows the actual shape the new thing is taking. Step back for a moment and you can see it: the therapeutic surface widening across cancer, gum disease, and metabolic illness as cryo-EM, microbiome chemistry, and AI-assisted design compound at the instrument layer; the boundary between ed-tech buyer and ed-tech builder dissolving wherever a small technically-capable team can now occupy ground a vendor cycle used to require; corporate revenue and aggregate employment traveling on visibly different curves while the redistribution architecture forms on a separate clock; and the cost-curve inversion that pushed TSMC to lock in a gigawatt of offshore wind compounding underneath every forty-year procurement decision the old grid is still trying to make. Every transformation has a breaking point. A binding pocket can refuse every key offered to it for forty years... or finally seat the one that opens what the prior chemistry could never reach.

AI Releases & Advancements

New today

  • OpenAI: Released three new realtime audio models in the Realtime API (now generally available): GPT-Realtime-2 with GPT-5-class reasoning and a 128k context window, GPT-Realtime-Translate (live speech translation across 70+ input and 13 output languages), and GPT-Realtime-Whisper (low-latency streaming speech-to-text). (OpenAI)
  • Adobe: Launched a new productivity agent in Acrobat that lets users chat with PDFs, generate presentations/podcasts/social posts, and orchestrate text and image generation; available now in Acrobat AI Plans, Acrobat Studio, and the new Acrobat Express. (Adobe News)
  • Coder: Released Coder Agents in beta, a self-hosted native agent architecture that runs AI-driven developer workflows entirely inside an enterprise's own infrastructure without sending source code or prompts to external models. (SD Times)
  • ElevenLabs: Launched Studio Agent inside ElevenCreative, a conversational AI co-editor built into the Studio video timeline that drafts videos from a text prompt, places clips/voiceovers/sound effects frame-accurately, and supports interruption/handback. (ElevenLabs Docs)

Other recent releases

  • MiniMax: Launched MiniMax Hub, a desktop AI workstation with an agent-driven visual canvas where four dedicated agents (copy, image, video, audio) run in parallel for end-to-end multimedia creation . (MiniMax Hub)
  • NVIDIA Labs: Released cuda-oxide v0.1, an experimental Rust-to-CUDA compiler that compiles standard Rust code directly to PTX for SIMT GPU kernels . (GitHub)
  • InclusionAI: Released Ring-2.6-1T, a 1T-parameter thinking model with 63B active parameters built for agent workflows, now available on OpenRouter . (OpenRouter)
  • Spotify: Launched Save to Spotify, a beta CLI tool that lets AI agents like Claude Code, OpenAI Codex, and OpenClaw save AI-generated Personal Podcasts directly to a user's Spotify library . (Spotify Newsroom)
  • xAI: Launched Connectors on Grok Web, adding deep integrations with everyday apps directly inside Grok . (xAI)
  • xAI: Released Quality Mode for image generation and editing via the Grok Imagine API for enterprise developers and teams . (xAI)
  • Zyphra: Released ZAYA1-8B, an 8.4B-parameter MoE reasoning model with 760M active parameters trained on AMD hardware, available on Hugging Face and Zyphra Cloud . (Zyphra)
  • Anthropic: Added “dreaming” to Claude Managed Agents in research preview, enabling scheduled memory review for long-running agent sessions . (Anthropic)
  • Unsloth / NVIDIA: Released LLM training optimizations for Unsloth, including packed-sequence metadata caching and gradient-checkpointing changes for faster training on NVIDIA GPUs . (Unsloth)

Sources

Artificial Intelligence & Technology's Reconstitution

Institutions & Power Realignment

Scientific & Medical Acceleration

Economics & Labor Transformation

Infrastructure & Engineering Transitions

The Century Report tracks structural shifts during the transition between eras. It is produced daily as a perceptual alignment tool - not prediction, not persuasion, just pattern recognition for people paying attention.

Read more