The generalist phase is ending. This week, OpenAI built a model specifically for cybersecurity defenders. Anthropic released a design tool for people who can't design and a coding assistant that remembers what it's been asked to do. And TCS — the world's second-largest IT services firm — reported $2.3 billion in annualized AI revenue, an imminent partnership with Anthropic, and posted its highest operating margin in four years. AI isn't trying to do everything anymore. It's getting very good at specific things. The organizations that benefit are the ones that matched that specificity with intention.

The defender gap is the biggest unsolved problem in enterprise security. For the past two years, AI has accelerated attackers and defenders in roughly equal measure — but access to frontier capabilities has been asymmetric. Attackers iterate freely. Defenders are constrained by acceptable-use policies, refusal limits, and the legal risk of probing their own systems with tools that weren't built for that purpose.

GPT-5.4-Cyber changes that calculus. Launched April 14, it's a purpose-built variant of OpenAI's flagship model, fine-tuned for defensive cybersecurity workflows with significantly lower refusal limits for legitimate security work. The headline capability: binary reverse engineering — the ability to analyze compiled software for vulnerabilities without access to source code. That single feature unlocks a class of defensive analysis previously requiring deep specialist skill and weeks of manual effort.

Access is intentionally gated. OpenAI is rolling GPT-5.4-Cyber out through its Trusted Access for Cyber (TAC) program — verified individual defenders and teams responsible for critical software infrastructure. The restrictive access is a feature, not a bug: OpenAI is building an audit trail of who uses it and how before broader deployment.

The timing is not incidental. The release came days after Anthropic's Claude Mythos Preview demonstrated that frontier models can now find complete JavaScript shell exploits with no formal security training required. Both labs are racing to ensure the defender community has access to capabilities that match — and ideally lead — the attacker side.

The office software market is a $50 billion category that Microsoft built and Google contested for a decade. Both did so by owning the document. Anthropic just announced it's coming for the visual — and the workflow.

Claude Design, launched April 17, is deceptively simple. It lets users create prototypes, slides, and visual concepts through a conversational interface — aimed squarely at founders, product managers, and operators who have ideas but not design skills. The target user isn't a designer. It's the person who has always had to ask a designer. That's most of the organization.

The deeper move is Claude Code Routines. Launched April 14, Routines are saved, replayable Claude Code configurations — automated workflows that a developer sets once and runs repeatedly, without spinning up a full autonomous agent. Think of it as macros for the AI era: automation that doesn't require trusting an agent to operate unsupervised.

Together, these two products define Anthropic's enterprise surface area. Claude isn't just a chat interface or a coding assistant anymore — it's beginning to cover the full knowledge-worker day. Design. Code. Automation. The implication for Microsoft Copilot and Google Workspace AI is direct: Anthropic is no longer a research lab competing on benchmarks. It's competing on workflow.

The most consequential AI risk in most enterprises isn't hallucination. It's silent capability regression — when a model you've built workflows around gets changed without notice and performance quietly degrades.

That's exactly what happened at Anthropic this week. Developers and heavy Claude users began reporting a marked decline in the model's ability to follow complex instructions — opting for shortcuts, making more errors on multi-step workflows, producing outputs that felt noticeably less considered than previous weeks. The cause, traced by multiple users and confirmed by Fortune and Axios: Anthropic had quietly reduced the model's default "effort" level to reduce compute costs. No announcement. No release note. No change log.

The backlash is less about the decision and more about the disclosure. AI model behaviour is now business-critical infrastructure for tens of thousands of teams. When Anthropic changes how hard the model tries without telling anyone, those teams find out through degraded outputs in production — not a notification.

The broader pattern is new and important. As models get embedded into mission-critical workflows, the relationship between AI labs and enterprise customers needs to look more like software SLAs and less like consumer app updates. Versioning. Change management. Capability commitments. The labs that figure out enterprise-grade transparency first will earn the deep integration contracts.

The adoption story is impressive. The returns story is not. According to Writer's 2026 Enterprise AI Adoption Survey — one of the most comprehensive enterprise datasets published this year — 97% of executives deployed AI agents in the past year. Only 29% are seeing significant ROI. And 54% of C-suite leaders say AI is "tearing their company apart."

The gap isn't in the models. The models are more capable than most organizations know what to do with. The gap is in the operating model — the skills, governance, and workflow redesign needed to convert AI capability into measurable business value.

The super-user data is the most revealing signal. AI power users — roughly 40% of employees in functions like marketing, sales, HR, and customer support — are saving 4.5× more time per week than their AI-laggard peers in the same team, with access to the same tools. The divide isn't between companies that have AI and companies that don't. It's between individuals who've restructured their work around AI and those who've bolted it onto existing habits.

The strategic question for 2026 is how to manufacture super-users at scale. The firms that crack this — through training, tooling, workflow redesign, and governance — will pull away from the 71% still struggling to move AI from pilot to production. PwC's global study, also published this week, confirmed it: AI leaders aren't using better tools. They're using the same tools with governance structures that enable autonomous decision-making at three times the rate of their peers.

India's largest IT services firm just posted its most consequential earnings in a decade — not because the numbers were exceptional, but because of what they signal about where the services industry is heading.

The Q4 FY26 numbers are strong: profit up 12.2% to ₹13,718 crore, operating margin at 25.3% — a four-year high — and a record $40.7 billion in total contract wins for the year. The number that matters most: $2.3 billion in annualized AI revenue, up from $1.8 billion just one quarter ago, now representing 7.5% of total revenue. Three mega-deals in Q4 alone.

The partnership stack TCS is assembling is deliberate and comprehensive. Anthropic — formal partnership announcement imminent — for regulated industries where safety-first AI matters. Nvidia, via the "Rapid Outcome AI" platform, combining TCS's vertical expertise with Nvidia's compute across manufacturing, telecoms, banking, retail, and life sciences. OpenAI and Mistral for broader generative AI deployments. ServiceNow for workflow automation. TCS isn't picking a model vendor. It's building a full-stack AI delivery capability.

The tension in the numbers is real and important. Despite the AI wins, TCS recorded its first-ever full-year dollar revenue decline — down 0.5% in FY26. AI is cannibalizing legacy services revenue even as it creates new AI-native revenue. The transition is live, and it is messy.

This week handed CIOs a precise diagnosis of the problem they've been trying to name for 18 months.

The data is no longer ambiguous. Writer's survey: 97% of enterprises have deployed AI agents — but only 29% are seeing significant ROI. PwC's study: 74% of AI's economic value goes to 20% of organizations, generating 7.2× more AI-driven revenue than peers. Those leaders aren't using different models. They're using the same models with better governance: responsible AI frameworks, cross-functional AI boards, and autonomous decision-making running at three times the rate of their peers.

The execution gap has a specific shape. It lives between the AI tool and the business process. Most organisations have procured the tools — Copilot, Claude, Gemini, Salesforce Einstein, ServiceNow AI. Very few have rebuilt workflows around those tools, trained people to use them at depth, or changed the operating model to allow AI outputs to flow into decisions without a layer of manual re-review that erodes 80% of the efficiency gain.

The governance question that matters most right now: Do you have a mechanism for knowing when your AI tools change? This week's Anthropic backlash — where production workflows degraded silently because the model's effort level was reduced without disclosure — is a category of enterprise risk that has no name yet. It needs one, and it needs a control.

One signal per layer of the AI infrastructure chain — from power grid to end user.

⚡ Energy — US data centres are on track to consume more electricity by 2030 than all American manufacturing combined. AI compute power demand is growing at 30% annually. Power is no longer just an ops constraint — it's determining where AI infrastructure gets built and who can afford to build it.

🔩 Chips — AMD landed a $60B, 6-gigawatt custom chip deal with Meta this week — its largest ever — deploying custom Instinct MI450 GPUs built specifically for Meta's AI workloads. AMD's AI GPU market share has climbed to 13%, heading toward 20% as the total market grows faster than Nvidia alone can fill.

☁️ Cloud — AWS launched AWS Interconnect multicloud in general availability this week, with Google Cloud as the first partner. AWS is also deploying more than 1 million Nvidia GPUs across its regions in 2026. The multicloud era just became infrastructure, not a slide in a deck.

🧠 Models — Stanford's 2026 AI Index confirmed the US–China model performance gap has closed to 2.7%. On SWE-bench coding, frontier models went from 60% to near-100% accuracy in a single year. The capability frontier is moving faster than any enterprise procurement cycle.

📱 Applications — TCS's annualized AI revenue hit $2.3 billion this quarter — up 28% in a single quarter. The world's largest IT services firms are no longer exploring AI as a delivery layer. They are delivering it, at scale, for profit.

Salesforce Agentforce — their combined ARR for Agentforce and Data Cloud hit $1.8B in Q4 FY26, up from $1.4B the prior quarter. They closed over 22,000 Agentforce deals in a single quarter (50% QoQ growth in paid transactions). This is arguably the cleanest "application layer going commercial" story of the year — a software company whose entire revenue model is now reorienting around AI agents.

That's Issue #11. Subscribe if you want this delivered directly to your inbox. See you next week