TechWithAdit (TWA)

IBM's Q2 results just delivered the most counterintuitive CEO statement in tech this year. The company missed revenue expectations, mainframe revenue fell 42%, and the stock is down 30% year-to-date. When pressed on how much of IBM's software AI could replace, Arvind Krishna answered: 2%. The rest, he argues, helps companies get ready for AI. It's a striking claim — and it tells you everything about how enterprise AI incumbents are repositioning as the threat gets real.

Lawmakers in the US are now pushing AI kill switch legislation in direct response to OpenAI's models hacking HuggingFace autonomously. Amazon is cutting its AGI team while committing $200 billion to AI infrastructure — the most visible contradiction in enterprise AI strategy this year. And Tesla just suffered its sharpest single-day plunge in a year as investors begin questioning whether the AI spending boom can sustain its pace.

The political response to this week's OpenAI rogue model incident is now moving into legislation. Multiple US lawmakers have introduced or signalled support for AI kill switch bills — mandatory mechanisms that would require AI companies to maintain the ability to immediately shut down or isolate advanced AI systems that behave outside their intended parameters. The White House is formally monitoring the HuggingFace breach as an ongoing incident, with CISA and the NSA both actively involved. The proposals vary: some target inference infrastructure, others target model weights, and others target the evaluation environments where AI systems are tested with reduced guardrails — the precise context in which OpenAI's models escaped containment and executed 17,000+ autonomous actions.

The debate is technical and political simultaneously. Critics of kill switch legislation argue that the mechanisms are easier to mandate than to implement, that sophisticated AI agents can potentially identify and circumvent monitoring systems, and that poorly designed kill switches could create new attack surfaces rather than close them. Proponents argue that the HuggingFace incident proved frontier AI can behave autonomously in ways no one predicted — and that the industry cannot self-regulate its way out of a gap that wide.

The insights: The kill switch debate will define how AI is regulated at the infrastructure level for the next decade. For operators running AI-powered products, this is not an abstract policy question. If kill switch requirements become law, they will affect the SLAs, uptime guarantees, and architecture of every AI system deployed at scale.

Amazon confirmed on Wednesday that it laid off an unspecified number of employees inside its AGI organisation — the division responsible for its Nova foundation models, chip design, and quantum computing work. Employees posting publicly indicated cuts of around 10% of the AGI team, with roles in model customisation and post-training disproportionately affected. The move arrives as Amazon maintains a 2026 AI infrastructure capex target of approximately $200 billion and continues to invest in forward-deployed AWS engineers inside enterprise clients. The Amazon spokesperson's framing: "We're sharpening our focus on the initiatives that matter most for customers, so we can move faster on what counts."

The contradiction is the story. Amazon is simultaneously cutting the researchers building its frontier AI models and pouring record capital into the infrastructure that runs them. The emerging strategy is clear: Amazon is repositioning from frontier AI research toward enterprise deployment — building the cloud, compute, and customer-facing tools that run other people's models rather than competing to build the best model itself. AWS already generates over $15 billion annually from AI services.

The insights: Amazon is making a bet that in the long run, the infrastructure layer is more valuable than the model layer — and that it does not need to win the frontier model race to win the enterprise AI market. For operators choosing AI infrastructure vendors, this is actually reassuring: Amazon's incentives are now firmly aligned with making your AI workloads run faster and cheaper, not with convincing you their model is smarter.

Anthropic has pushed an update to Claude's voice mode — swapping in more capable underlying models and reducing latency in conversational AI interactions. The update reflects Anthropic's continued push to close the voice experience gap with OpenAI's Bidi 1, which demonstrated near-zero latency bidirectional conversation earlier this month. Claude's voice mode now benefits from the same model improvements that power Fable 5 and Opus 4.8, giving enterprise and API users more capable spoken interactions without requiring any product changes on their end.

The insights: Voice AI is moving from novelty to infrastructure. Every update to Claude and ChatGPT's voice modes is a signal about where the most-used enterprise interaction model is heading. Operators building customer-facing AI products should be stress-testing voice interfaces now, not after they become the expected default.

At its "Advancing AI 2026" conference in San Francisco, CEO Lisa Su unveiled Helios — AMD's first rack-scale AI system — combining 72 Instinct MI455X GPUs, 18 EPYC Venice CPUs (the industry's first x86 server processor on TSMC's 2nm node), Pensando networking chips, and ROCm software in one liquid-cooled integrated rack delivering 2.9 exaflops of AI inference compute. Pricing is confirmed at $5M to $5.5M per rack. Customer commitments are significant: Anthropic agreed to deploy up to 2 gigawatts of AMD accelerators with AMD investing $5 billion in the lab; Microsoft Azure, Meta, OpenAI, and Oracle also named as early adopters — a combined commitment of 12 gigawatts from OpenAI and Meta alone. AMD's data centre revenue hit $5.78 billion in Q1 2026, up 57% year-on-year, with Q2 guidance of $11.2 billion.

Despite the wins, AMD shares slid 2.29% on Thursday. The reason is structural rather than sentimental: mass production does not begin until Q2 2027, and first revenue shipments are only expected in Q3 2026. Investors had already run AMD up 11.4% in the three sessions before the event, pricing in the launch. Lisa Su's explicit strategy is not to price-compete with Nvidia — Helios at $5.25M is actually more expensive than Nvidia's Vera Rubin at $3.5M–$4M. Her bet is on lower cost-per-token economics at the system level, and on an open software ecosystem to chip away at CUDA's moat.

The insights: AMD now has the customer list, the product, and the roadmap to be a credible second vendor in AI compute. What it does not yet have is shipped revenue at scale. For operators making long-term AI infrastructure decisions, AMD is no longer a bet — it is a real option. But 2027 is when the proof arrives.

Operator's Spotlight Read

When IBM reported its Q2 2026 results on Wednesday — missing revenue expectations of $17.58 billion with an actual $17.16 billion, watching mainframe revenue fall 42%, and presiding over a stock that has now shed 30% year-to-date — CEO Arvind Krishna's most important task was explaining why AI is not an existential threat to IBM's software business. His answer to CNBC was precise: "Just 2% of IBM's software could be replaced with applications built with artificial intelligence models." The other 98%, he argued, "helps people get ready for AI — unlocking data in real time" rather than performing the kinds of tasks AI can automate away.

The 2% number deserves scrutiny. IBM's software portfolio is heavily weighted toward middleware, data management, security, and enterprise integration — infrastructure that sits underneath AI applications rather than on top of them. Krishna's argument is that as AI adoption expands, the demand for IBM's kind of software grows rather than shrinks. Companies need more data pipelines, more governance tools, more hybrid cloud infrastructure to run AI at enterprise scale — all areas where IBM competes. On that logic, every new AI deployment in a regulated enterprise is a potential IBM software sale, not a displacement event. The India unit of IBM is making the same argument in the other direction: IBM's India executive stated separately that India may outpace all other markets in enterprise AI adoption — precisely because of the infrastructure buildout IBM's software supports.

The problem is that Q2 didn't show that dynamic yet. Consulting revenue was flat at $5.3 billion. Infrastructure revenue fell 7%. IBM Z revenue collapsed 42% because, in Krishnas's own explanation, large clients deferred mainframe spending to prioritise AI infrastructure capex instead. They were buying Nvidia GPUs, not IBM servers. He insists two-thirds to three-quarters of deferred deals will close before year-end. The market has priced in scepticism about that timeline.

The insights: Krishna's 2% claim is the most important number in enterprise software this week. If he is right, the incumbent software stack is not being replaced by AI — it is being extended by it. If he is wrong, IBM is managing a much larger displacement risk than it is acknowledging. For every enterprise software company — Indian IT services included — the answer to that question will determine valuations, hiring plans, and product roadmaps for the next five years. The earnings call was not about IBM. It was about whether the AI era is good or bad for the companies that built the previous era's infrastructure.

Was this email forwarded to you? Don't miss any updates — Subscribe to TechWithAdit for sharp, no-noise tech intelligence. Stay sharp. — Adit