The week closes with a credible challenge to Nvidia's software moat, a newly mapped attack surface inside enterprise AI deployments, and workforce data that complicates the displacement narrative. None of these developments resolves overnight, but each changes the calculus for operators making infrastructure, security, and hiring decisions.
Qualcomm places a $14 billion bet on a CUDA-free stack
On 24 June, Qualcomm announced it would acquire Modular — the AI infrastructure company founded by LLVM and Swift creator Chris Lattner — in an all-stock deal valued at $3.9 billion. Modular built Mojo, a programming language designed to replace Python for AI workloads, and MAX, an inference engine that compiles once and runs across Nvidia, AMD, Intel, Apple Silicon, and Qualcomm hardware without rewrites. At the same Investor Day, Qualcomm raised its FY 2029 non-handset revenue target to $40 billion and named data-centre AI inference as the primary driver.
Separately, Qualcomm is in advanced negotiations to acquire Tenstorrent — Jim Keller's RISC-V AI chip startup — at a valuation between $8 billion and $10 billion. Tenstorrent's Blackhole chip uses a RISC-V mesh architecture in which each compute unit bundles five processors with dedicated matrix and vector engines. If both deals close, Qualcomm will have assembled a full alternative to the Nvidia stack: open-ISA silicon paired with a hardware-agnostic compiler, at a combined outlay approaching $14 billion.
CUDA's durability is not the chip; it is the ecosystem of libraries, tooling, and institutional habit built on top of it. The Modular deal directly attacks that layer. The strategy will not dislodge Nvidia in the near term, but it is the first capital commitment at this scale targeting the software moat rather than just the hardware. Operators locked into CUDA-specific workflows should now be asking their vendors what a migration path looks like.
Prompt injection is a structural vulnerability, not a bug
A VentureBeat analysis published on 28 June synthesises the year's enterprise AI security picture and arrives at an uncomfortable conclusion: prompt injection is structural. The same architecture that gives an agent the ability to act — reading from RAG pipelines, routing across models, writing to corporate systems — is the architecture attackers are now exploiting systematically.
The underlying data comes from the CrowdStrike 2026 Global Threat Report, which documented adversaries exploiting generative AI tools at more than 90 organisations in 2025, using injected prompts to generate commands that stole credentials and moved laterally. The average eCrime breakout time fell to 29 minutes; the fastest recorded was 27 seconds. New attack vectors identified in 2026 include RAG poisoning — seeding malicious instructions into documentation or GitHub repositories that get ingested into enterprise knowledge bases — and router manipulation, where attackers force a pipeline to switch to a weaker model to bypass guardrails.
The practical implication for operators deploying agents that write emails, execute code, or modify cloud infrastructure: every external data source feeding that agent is a potential attack vector. Defences that live inside the model's own reasoning layer are insufficient; enforcement must sit at the infrastructure level, below the agent's decision space. Treating all ingested external content as potentially hostile is not paranoia — it is the architecture.
Engineers are holding — but the entry-level pipeline is in crisis
SignalFire's State of Talent Report 2026, drawn from hiring data across 80 million companies, provides the clearest picture yet of AI's workforce impact at large technology firms. Engineers accounted for 55 per cent of all new hires in 2025 across twelve Tech Majors including Alphabet, Meta, Apple, Amazon, and Microsoft — up from 46 per cent in 2019. While total hiring at these companies fell 25 per cent against 2019 levels, engineering roles declined by only 11 per cent. The pattern holds at early-stage startups, which hired 7 per cent more engineers in 2025 than in 2019.
The headline is not uniformly good news. Entry-level engineering hiring at Tech Majors is down 65 per cent versus 2019, and new-graduate hiring across the sector has fallen 50 per cent. The pattern that emerges is a compression toward a senior-led engineering core: AI tools amplify individual output enough that fewer, more experienced engineers can cover the same surface area, while the traditional apprenticeship pathway for early-career developers is being dismantled in real time.
For operators making headcount decisions, the data points to a bifurcation risk. Senior engineers are in short supply and rising demand. A generation of potential engineers cannot get a first role. Teams that eliminate junior positions to cut costs today may face a narrower pipeline of experienced talent three to five years from now.
OpenAI refreshes GPT-5.5 Instant with sharper intent inference
On 24 June, OpenAI pushed an updated version of GPT-5.5 Instant to ChatGPT and the API. The update focuses on conversational quality: the model is now more reliable at inferring the underlying goal behind a multi-part question, handling requests with several competing constraints, and producing shopping and local-service recommendations that account for full context. Pricing is unchanged; the update reached free-tier users by 25 June and is accessible via the chat-latest API alias.
The relevance for enterprise operators running production agentic workflows is limited — GPT-5.5 remains the appropriate alias for stable production use. But GPT-5.5 Instant is the model behind the majority of consumer ChatGPT interactions, and the intent-inference improvements make multi-step conversational flows — customer support, internal helpdesks, decision-support tools — noticeably more coherent at no additional cost.
Taken together, this week's developments describe a maturing competitive landscape. Infrastructure is attracting serious challenger capital for the first time. The agentic security attack surface is being mapped rather than ignored. Workforce data is granular enough to inform real hiring decisions. The decisions operators defer this quarter compound into next year's constraints.