OpenAI’s infrastructure moves: Broadcom and AMD partnerships

AMD partnership — securing GPU capacity

• OpenAI and AMD recently announced a 6 gigawatt multi‑generation compute agreement. 

  • The first 1 GW deployment is expected in H2 2026, using the AMD Instinct MI450 GPUs.

  • To align incentives, AMD granted OpenAI warrants to acquire up to 160 million shares, vesting as deployments scale. 

  • This deepens their hardware‑software alignment and gives OpenAI more predictability over supply.

• The move also signals a shift away from exclusive dependence on one vendor (e.g. Nvidia) and adds resilience to OpenAI’s supply chain.

• That said, execution risk is nontrivial (manufacturing yields, integration, cooling/power, orchestration across heterogeneous hardware) — it’s a big bet, not a guarantee.

Broadcom partnership — custom AI accelerators

• OpenAI also announced a strategic collaboration with Broadcom to co‑develop 10 gigawatts of custom AI accelerators + network systems.

  • OpenAI will handle the design side (model‑aware optimizations, embedding their algorithmic insights), while Broadcom will build and deploy the systems. 

  • The rollouts are scheduled from H2 2026 through end of 2029. 

  • The systems will include Broadcom’s Ethernet, PCIe, optical interconnects to support scale-out and scale-up networking.

• The rationale: by co‑designing hardware that’s tightly aligned with their models and workloads, OpenAI can extract performance, latency, and power gains that are hard to get from commodity chips. Also, it gives them more control over cost structure, supply chain, and differentiation.

• It’s also a hedge: if GPU vendors become constrained (due to demand, export controls, etc.), having their own custom “accelerator fleet” gives OpenAI more autonomy.

Strategic considerations & risks

• Heterogeneity & software abstraction: Running efficiently across multiple hardware types (AMD GPUs, Broadcom accelerators, maybe third‑party hardware) demands robust abstraction layers, compilers, memory & interconnect strategies.

• Scale & operations: Deploying on the GW scale means huge demands on power, cooling, reliability, and infrastructure management.

• Lock-in vs openness: The more custom the hardware, the harder it is for others to replicate (good for defensibility), but also harder to maintain ecosystem compatibility and standard tooling.

• Technology risk & obsolescence: By the time some of these systems come online (2028–2029), model architectures might shift; hardware must be adaptable.

Overall, this combination of a major GPU commitment and in-house custom accelerator design signals that OpenAI is doubling down on owning the full stack — from model to hardware — to sustain performance leadership and margin control.

Codex, ChatGPT apps, and the evolving AI agent ecosystem

Codex becomes Generally Available (GA)

• OpenAI announced that Codex is now generally available (no longer just a preview).

  • New features include a Slack integration (you can @Codex in a Slack thread, and it will gather context and respond) 

  • They also released a Codex SDK, enabling embedding the Codex agent capabilities into your own tools and applications. 

  • For workspace admins, OpenAI added environment controls, monitoring, and analytics features to manage usage at scale.

  • The underlying model is GPT‑5‑Codex, a version of GPT‑5 optimized for agentic coding. 

• In the preview period, Codex’s usage grew rapidly (10× in 2 months).

• The adoption inside OpenAI was also strong: nearly all internal engineers are now using it, merging ~70% more pull requests weekly than earlier. 

This is a shift: Codex is no longer a side experiment — it’s entering “production” status, with tooling, access, analytics, and embedding capabilities.

Apps in ChatGPT + Apps SDK

• OpenAI also introduced “apps in ChatGPT” along with an Apps SDK (preview). OpenAI

  • Developers can build interactive apps that live inside ChatGPT — so when a user is conversing, an app might “activate” contextually to help. 

  • Examples: you might ask ChatGPT to design a poster (via Canva), then follow up with creating a pitch deck. The apps handle the specialized UI. 

  • Apps respond to natural language and include interactive UI elements in the chat context. 

  • Privacy & permission controls are built in: when an app is first connected, ChatGPT prompts you on what data is shared. 

  • The roadmap is to support apps in ChatGPT Business, Enterprise, and Edu, and open a directory for users to discover apps. 

• This effectively turns ChatGPT from “just a chatbot” into a platform, a container for third-party services and UIs — much like how smartphone OSs host apps.

• For users, this means more seamless workflows: you don’t need to exit the chat to call specialized services. For developers, it's a new channel for embedding capabilities directly into conversations.

• This also reflects a broader trend: “agentic” systems — AI that can call tools, fetch data, execute code — are becoming foundational, not edge features.

Other changes & enablers

• Under the hood, updates to Codex include faster, more reliable performance, better real-time collaboration, and integration across terminal, IDE, web, and even phones. Codex’s unified experience lets you hand off between local editing and cloud tasks without losing context. 

• Importantly, Codex cloud tasks will soon count against your usage quota (from Oct 20 onward). 

Takeaway: OpenAI is pushing hard to position Codex not as a toy for demos, but as a full-fledged “coding coworker” — and to make ChatGPT itself a canvas for interacting with domain-specific services (apps)

Google’s advances: Veo 3 and Gemini Robotics / Embodied AI

Google’s video tools: Veo 3 / Flow updates

• Google is enhancing its AI-powered video editing/generation tool (Flow / Veo 3). 

• In the 3.1 update, users can modify lighting and shadows in generated video, making the output more realistic. 

• New “Ingredients to Video” allows video + audio generation from three reference images. 

• “Frames to Video” can transition between a start and end image, generating intermediate frames + sound. 

• “Scene Extension” lets you expand a video clip by up to 1 minute with AI-generated visuals and audio. 

• Object removal is coming: remove an item, and the system reconstructs the scene to eliminate traces. 

• Veo 3.1 is currently in “paid preview” via the Gemini API. 

• These features push AI video closer to full creative editing — not just content generation but tool-like capabilities to refine, extend, and manipulate visuals.

• This is part of a broader push (across Google and rivals) to integrate generative video into everyday creation flows (production, marketing, creative content).

Gemini Robotics / Embodied AI

• Google DeepMind introduced Gemini Robotics, a family of AI models that bridge the gap between digital reasoning and physical action. 

  • The models are built on top of Gemini 2.0 and include Vision-Language-Action (VLA) capabilities. 

  • They can process images, video, and natural language prompts to generate actions for robots — e.g. manipulation, object interaction, movement in dynamic environments. 

  • The Gemini Robotics‑ER (Embodied Reasoning) variant extends multimodal reasoning into 3D spatial and temporal understanding: bounding boxes, trajectory planning, multi-view correspondences. 

  • The Gemini Robotics‑ER 1.5 API lets robots “understand” a scene: detect objects, plan sub‑actions, and execute them by invoking functions or code.

• The ambition: robots that can interpret natural language commands such as “put the apple in the bowl,” break them down into sub-actions (pick, move, place), and carry them out in real environments.

• The shift is toward generalist robotics — not rigid, narrowly programmed systems, but embodied agents that integrate perception, reasoning, and action.

• Challenges remain: robustness in unstructured environments, safety, long-horizon planning, real-time adaptation, hardware constraints (sensing, actuation), and sim-to-real transfer.

---

Broader trends: AI in science, enterprise uptake, etc.

AI in scientific discovery

• Research continues to push AI toward assisting in advanced scientific tasks: hypothesis generation, data interpretation, experiment planning, and cross-domain integration.

• For example, a new benchmark SciVideoBench aims to assess video reasoning in scientific domains, where models must interpret experimental videos and answer domain-specific questions.

  • Early evaluations show current state-of-the-art models (even proprietary ones) struggle significantly in these tasks, indicating much headroom. 

• Also, embodied AI models (like Gemini Robotics) and multimodal video models (e.g. “AlanaVLM” for egocentric video reasoning) are beginning to bridge perception, action, and scientific contexts. 

• The interplay between AI and science is accelerating: AI tools help interpret high-throughput experiments (e.g., in genomics, high-energy physics, material science) more rapidly, enabling new hypotheses and insights.

• But caution: scientific domains require strong rigor, domain validity, explainability, and safety. The “hallucination” or overconfidence of AI remains a risk. The models must be grounded, verifiable, and interpretable.

Enterprise AI adoption and momentum

• Across industries, enterprises are increasingly embedding AI models into workflows, products, and decision systems — not just as experimental pilots, but as core infrastructure.

• Key enablers:

  • Scalable infrastructure: the compute deals (AMD, Broadcom) and cloud expansion make it more feasible to host large models in enterprise settings reliably.

  • Tooling & integration: SDKs, admin controls, embedding agents (Codex, apps in ChatGPT) reduce the friction of integrating AI into existing systems.

  • Vendor partnerships: e.g. OpenAI’s expanded alliance with Salesforce to integrate ChatGPT + their models into the Salesforce ecosystem. 

  • New business models: AI-powered assistants, agentic tooling, “AI as a service” embedded in traditional products.

• But there are still obstacles: data privacy, interpretability, alignment with domain constraints, cost (compute and licensing), regulatory & compliance issues, and workforce adaptation.

• The shifts in infrastructure also raise barriers to entry: smaller teams or companies may struggle to compete if they can’t access large-scale compute, custom hardware, or high-fidelity agents.

---

Synthesis and forward look

Putting all of this together, here’s how I see the trajectory:

1. Full-stack control is becoming essential

   • It’s no longer sufficient to rely purely on third-party GPUs; leading AI players are working to own every layer (model, hardware, orchestration). The OpenAI + AMD + Broadcom moves are emblematic of that.

2. ChatGPT evolves into a platform & agent hub

   • With apps inside ChatGPT and embedding SDKs, the “chat interface” becomes the universal shell. You won’t need to leave the conversation to call services — they’ll be embedded and context-aware.

3. Agentic AI is entering the mainstream

   • Tools like Codex, AgentKit (drag‑and-drop agents), app SDKs, robotics agents (Gemini Robotics) all point to a world where AI systems will actively carry out tasks for users, rather than just respond.

4. Perception → reasoning → action bridging

   • The push into embodied AI and robotics (e.g. Gemini Robotics) shows the ambition to connect “thinking AI” with the physical world, closing the loop between perception, planning, and execution.

5. Scientific and industrial AI rise

   • As models mature, more of their impact will come in serious domains: scientific discovery, industrial automation, healthcare, logistics, etc. The barrier is less about “can AI do this” and more about “can it do it reliably, verifiably, safely.”

6. Consolidation and scale economics

   • The compute and infrastructure demands will favor large players. The cost of entry is rising. Smaller entities will need to differentiate via niche specialization, algorithmic innovation, or partnering.

7. Governance, safety, and alignment will increasingly matter

   • As AI agents act in the world and have more autonomy, issues of safety, alignment, auditability, verification, adversarial robustness, and control are becoming urgent, not academic. 

   • complied by 

   • aqsa mahak( financial analyst)

---

Major AI Market Trend: Infrastructure & Compute Arms Race

   Major AI Market 

 Trend: Infrastructure & Compute Arms Race

What’s happening

• The demand for compute power, specialized hardware, and high‑performance infrastructure is growing explosively. AI models (especially large language models, generative AI, multimodal models) need massive computation, memory, and fast interconnects.

• AI supercomputers and data centers are becoming critical bottlenecks. One recent study shows AI supercomputer performance has roughly doubled every 9 months (though cost and power also double annually). arXiv

• Hyperscalers (the big cloud providers) are increasing capital expenditures (CapEx) heavily toward AI infrastructure—servers, chips, networking, cooling, power, data center expansion. Global X Japan+2StartUs Insights+2

• Hardware companies (GPUs, AI accelerators, memory, networking) are reaping the benefits. For instance, demand for AI servers has pushed firms like Dell to raise growth forecasts. Reuters

• The AI market as a whole is projected to grow rapidly: from ~$371.7 billion in 2025 to over $2,400 billion by 2032 (CAGR ~30 %) per some market reports. StartUs Insights+2Global X Japan+

 

Thus, among the many AI subtrends (software, applications, services), infrastructure / compute is arguably the engine fueling the r

---

Why it matters

• Strong moat / barrier to entry: Building high performance hardware and data center infrastructure is capital intensive. New entrants struggle to compete at scale.

• Supply constraints: Scarcity of next‑generation chips, yield challenges, and lead times can delay deployment, giving incumbents leverage.

• Ecosystem leverage: Infrastructure firms often partner with cloud providers, enterprises, governments. Being part of the foundational layer gives optionality into higher layers (software, services).

• Predictability: Infrastructure demand tends to have longer visibility (multi‑year purchase cycles), making it more “investable” relative to speculative application bets.

---

Risks & potential turning points

• Valuation overshoot: Many infrastructure / AI hardware plays are commanding rich valuations based on future potential rather than current earnings. Some academic work argues there’s a “valuation misalignment” where expectations exceed realized performance. 

• Technological discontinuities: A breakthrough in AI algorithm efficiency (requiring far less compute) or a new architecture (neuromorphic, quantum) could displace heavy compute reliance.

• Power / energy constraints: AI infrastructure is extremely power intensive. Rising energy costs, regulatory constraints, or ecological limits could slow expansion. The data shows that cost and power needs double roughly annually for these supercomputers. arXiv

• Geopolitical / supply chain risks: Concentration of chip manufacturing (e.g. in Taiwan, China, US) is vulnerable to trade wars, export controls, or conflict disruptions.

• Integration / utilization risk: Having infrastructure doesn’t guarantee that businesses will utilize it well. If enterprise adoption or AI monetization lags, returns may suffer.

complied by 

Aqsa mahak  (financial analyst )