!IBM shares plunge as Anthropic touts COBOL modernization
The Arc: A Blog Post Worth $31 Billion
IBM lost $31 billion in market cap today because of a blog post.
Not a product launch. Not a regulatory action. Not an earnings miss. A blog post, from Anthropic, announcing that Claude Code can automate the exploration and analysis phases that have made COBOL modernization prohibitively expensive for decades. IBM shares dropped 13.2%, the steepest single-day decline since the dot-com crash in October 2000. Accenture and Cognizant fell 6–7% in sympathy. The Dow lost over 800 points, dragged by IBM's weight in the price-weighted index.
Meanwhile, on the hardware side, two announcements landed that frame what comes after the current GPU era: OpenAI revealed that its first Jony Ive-designed device is a $200–300 smart speaker shipping early 2027, not a phone, not glasses, but an ambient AI presence with a camera and facial recognition. And a startup called Taalas unveiled the HC1, a chip that hardwires AI models into silicon and delivers 17,000 tokens per second, 10x faster than anything on the market.
The pattern across all three stories: AI is repricing the things we thought were permanent. Sixty-seven-year-old programming languages. The form factor of computing devices. The assumption that GPUs are the only way to run inference.
Monday was the day the market started processing what "disruption" actually means when AI moves from demo to deployment.
BLUF At Neue Alchemy, we support leaders navigating inflection points, when tech, capital, and policy converge. If your roadmap is already in motion and you're pressure-testing execution, we're open to conversations.
We also reserve capacity for education, SMBs, and mid-market leaders, those starting, mid-flight, or seeking outside perspective before systems harden.
MARKET & DISRUPTION Anthropic's COBOL blog post wiped $31B from IBM's market cap in a single session, the worst day in 25 years.
!IBM is the latest AI casualty, CNBC
Anthropic published a blog post Monday morning explaining how Claude Code can map dependencies across thousands of lines of COBOL, document workflows, trace execution paths, and surface risks that would take human analysts months to identify. The message was clear: "Legacy code modernization stalled for years because understanding legacy code cost more than rewriting it. AI flips that equation."
IBM shares fell 13.2% to $223.35, wiping $31 billion in market cap. February is now tracking a 27% monthly decline, on pace for IBM's worst month since at least 1968, per Bloomberg. The selloff dragged the Dow down over 800 points. Accenture and Cognizant dropped 6–7%, extending a broader pattern: every major Claude Code announcement in recent weeks has triggered sector-wide declines in legacy tech and services stocks.
IBM pushed back immediately, noting it has its own watsonx Code Assistant for Z and that the latest z17 mainframe cycle is outperforming its predecessor. Evercore ISI maintained its Outperform rating with a $345 target. Trefis called the selloff an overreaction.
So What? The market isn't pricing whether Claude Code can actually replace IBM's COBOL business today. It's pricing the narrative shift, the moment investors stopped treating legacy entrenchment as a permanent moat. IBM still controls 95% of ATM transaction infrastructure and generates strong mainframe revenue. But the perception that a startup's blog post can threaten that position changes how every legacy IT stock gets valued going forward. This is not a technology event. It's a repricing event.
The Risk: If you're an enterprise running COBOL on IBM mainframes, the risk isn't that Claude Code replaces your system tomorrow. It's that your board and CFO now believe it can, which changes every budget conversation about modernization timelines, consulting spend, and vendor lock-in. The "sell first, ask questions later" dynamic doesn't need to be technically correct to reshape procurement.
Action: If you're paying for COBOL modernization consulting, benchmark your current engagement against what Claude Code can actually do today, not what the blog post implies. The gap between AI-assisted analysis and full-stack modernization is still massive. But the pricing leverage just shifted permanently. If you're an investor, watch IBM's Q1 guidance for mainframe-specific revenue commentary.
HARDWARE & DEVICES OpenAI's first Jony Ive device is a $200–300 smart speaker with a camera, shipping early 2027.
!Jony Ive's first OpenAI device, MacRumors
Per The Information, OpenAI's 200+ person devices team, formed after the $6.5B acquisition of Ive's startup Io Products in May 2025, is building a smart speaker with persistent listening (no wake word), a camera for environmental awareness, and Face ID-style facial recognition for purchases. The internal pitch to employees described a device that "observes users and suggests actions to help them achieve goals."
Apple veterans lead the hardware effort: Tang Tan on hardware, Evans Hankey on industrial design, Scott Cannon on supply chain. Smart glasses and a smart lamp are also in development but won't arrive until 2028 or later.
Internally, tensions have surfaced between LoveFrom's design process and OpenAI's engineering team, LoveFrom has been slow to revise designs and shares little about its methodology, even with OpenAI's own devices staff.
So What? The product itself, a smart speaker with a camera, isn't novel. Amazon has been doing this for a decade. What's novel is the bet that conversational AI has gotten good enough to replace the app-and-screen paradigm entirely. OpenAI is wagering that by 2027, asking ChatGPT to buy something through a speaker will feel more natural than tapping through an app. If that bet is right, the speaker is the trojan horse for an AI-mediated commerce layer that bypasses Google, Apple, and Amazon's existing distribution.
The Risk: The hardware track record for software companies is brutal. Google, Facebook, and Amazon all stumbled before finding their footing (or abandoning the effort). OpenAI has never shipped a physical product. The Ive mystique buys attention, but attention doesn't survive a bad review cycle. And a camera-equipped, always-listening AI device enters a privacy conversation that's only getting louder.
Action: Don't plan around this device yet, it's 12+ months from shipping and the form factor could change. But do pay attention to the interaction model. If OpenAI successfully trains users to buy things through voice + visual AI, the implications for e-commerce, advertising, and brand discovery are significant. Start mapping what "AI-mediated purchase" means for your category now.
COMPUTE & SILICON Taalas unveiled HC1, a chip that hardwires AI models into silicon, 10x faster than anything on the market at 1/20th the cost.
Toronto-based Taalas emerged with the HC1, a custom chip fabricated on TSMC's 6nm process that permanently embeds Meta's Llama 3.1 8B model into the physical wiring of the silicon. The result: 17,000 tokens per second per user, roughly 10x faster than Cerebras and 100x faster than a standard GPU, at 1/10th the power and 1/20th the build cost. No HBM, no liquid cooling, no advanced packaging.
The startup raised $169M this round, bringing its total above $219M. The team is 24 core engineers, including alumni from AMD, Apple, Google, Nvidia, and Tenstorrent. Founder Ljubisa Bajic previously co-founded Tenstorrent and held architect roles at both AMD and Nvidia.
The catch: the HC1 only runs the model baked into it. Llama 3.1 8B is small and far from the frontier. But Taalas claims it can retool chips for new models in two months using a proprietary workflow with TSMC. A mid-sized reasoning model is expected this spring; a frontier model on second-gen silicon (HC2) by winter.
So What? Taalas is testing a radical thesis: that the era of general-purpose AI compute is a transitional phase, not an endpoint. If inference demand stabilizes around a few dominant models, and if the cost of minting model-specific silicon drops to weeks instead of years, then the GPU's flexibility premium becomes a tax, not an advantage. The economics are striking: 10x speed, 20x lower cost, 10x less power. If that scales to frontier models, it rewrites the entire inference cost curve.
The Risk: The model baked into HC1 is functionally a demo. Llama 3.1 8B is 18 months old and nowhere near competitive with Claude, GPT-4, or Gemini. If Taalas can't scale to frontier-class models by late 2026, the technology becomes a footnote. The "two months from weights to silicon" claim also hasn't been stress-tested at production scale.
Action: If you're running inference at scale, put Taalas on your evaluation roadmap for H2 2026 when the frontier silicon is expected. The approach is too early to adopt but too compelling to ignore. If you're investing in AI infrastructure, watch whether the HC2 generation validates the cost and performance claims at larger model sizes. This is the first credible challenge to the assumption that inference requires GPUs.
CONTRARIAN SIGNAL The most dangerous thing in AI isn't the model. It's the blog post.
A blog post moved $31 billion in market cap. Not a product. Not a benchmark. Not a customer win. A blog post that described a capability and let the market's imagination do the rest.
This is the new disruption pattern. The technical reality, that COBOL modernization still requires deep human expertise for strategy, compliance, testing, and integration, didn't matter. What mattered was the narrative: that AI can now do the thing that was supposed to be too hard to automate.
The contrarian read: the companies most at risk from AI aren't the ones whose technology gets replaced. They're the ones whose perceived moats get repriced. IBM's actual mainframe business may be fine. IBM's stock may not be, because investor confidence in the moat's permanence just cracked.
If your competitive advantage depends on the market believing something is too complex to automate, today was a warning shot.
The Takeaway: Moats built on complexity are only as durable as the market's belief in that complexity. AI doesn't have to replace you to reprice you.
THE QUESTION FOR TODAY
A blog post erased $31 billion.
A startup etched an AI model into silicon.
A tech CEO hired Jony Ive to build a speaker that watches you.
If the old assumptions about what's hard, what's fast, and what's permanent are all getting repriced simultaneously, what's the assumption your business is built on, and is it next?
See exactly how this impacts your specific industry and function. Upgrade to PRO to get bespoke tactical breakdowns generated instantly for your operating model.
Go deeper with the Weekly Signal
This is the daily take. The Weekly goes further — full strategic analysis across 8–10 sections, each with a signal read and operator action items. Source panel included.
Sign up free → then upgrade