Agentic AI Sounds Inevitable. Revenue Still Isn’t.

Today, we’re diving into:

Is Jensen Huang ringing the bell at the top of the AI bubble?

There are two ways to read Jensen Huang’s latest earnings call.

Nvidia posted numbers that would’ve melted screens in any other cycle: Q4 revenue up 73% to $68.1 billion, with guidance suggesting another 78% expansion. Yet the stock barely moved.

Why? Because markets are no longer asking “Is demand strong?”

They’re asking: How long can hyperscalers keep writing these checks?

More than half of Nvidia’s revenue comes from five cloud giants like Google, Amazon, and Meta, all racing to build AI data centers.

Collectively, they’re budgeting nearly $700 billion in capex this year. Meta alone plans up to $135 billion, Google up to $185 billion. That’s already stretching beyond free cash flow and into debt markets.

If that capex doubles annually, we’re staring at $2.8 trillion by 2028 and $5.6 trillion by 2029. That’s not incremental growth, that’s literally economic re-architecture.

Huang’s logic is simple, almost disarmingly so: the world used to spend $300–$400 billion annually on classical computing. AI requires 1,000x more computation.

If tokens are the new oil, then token-generation capacity must scale accordingly. Therefore, spending scales.

The real inflection, though, isn’t just larger models. It’s agentic AI.

When Huang says “Agentic AI has reached an inflection point,” he’s not talking about chatbots. He’s talking about autonomous systems that initiate actions, call tools, orchestrate workflows, and recursively generate more compute demand.

Agents don’t just answer queries, they also create new ones. They spin up simulations, trigger downstream models, and expand token consumption exponentially.

That’s the structural bull case: agents → more tokens → more GPUs → more data centers.

After that comes “physical AI” like robotics, manufacturing systems, industrial automation. Thus, compute moves from cloud prompts to real-world control loops.

So is this the top? Or the beginning of a multi-decade compute supercycle?

It’s actually a high-stakes infrastructure bet with asymmetric outcomes. The difference this time is concentration. In past bubbles, excess was distributed across thousands of weak balance sheets.

Today, the spending is concentrated among a few hyperscalers with deep cash reserves, diversified revenue streams, and strategic necessity driving their decisions. That makes the cycle more durable, but also more interdependent.

If agentic AI truly becomes embedded in enterprise workflows by autonomously negotiating contracts, optimizing logistics, writing production code, managing supply chains, then compute demand won’t be speculative. It will be operational.

Tokens then won’t just be consumed for experimentation; they’ll be tied directly to revenue-generating activity.

But here’s the nuance business owners should internalize: infrastructure cycles don’t automatically translate into startup wins. Massive capex at the platform layer often compresses margins at the application layer.

Thus, when compute becomes abundant, differentiation shifts to distribution, proprietary data, workflow integration, and execution speed.

The real opportunity here isn’t to mirror Big Tech’s spending.

It’s to position your company where expanding compute lowers your marginal cost or increases switching costs for customers. If this is a supercycle, the spoils won’t go to those who shout “AI” the loudest, they’ll go to those who build durable economic value on top of it.

Agentic AI is the new Silicon arms race

The launch of the Samsung Galaxy S26 signals something bigger than a hardware refresh. It marks the transition from phones that respond to you, to devices that act for you.

To understand the shift, you need to separate agents from agentic systems.

An AI agent, whether powered by Google Gemini, Perplexity, or an evolved Bixby operates within a bounded execution graph. It processes a prompt, executes a defined task, and returns output.

The control loop is external; you provide intent explicitly at every step. Technically, this is event-driven inference with stateless or lightly stateful memory constraints.

Agentic AI introduces a persistent reasoning loop.

At the chipset level, enabled by processors like the Snapdragon 8 Elite Gen 5, this means continuous context ingestion across application states, sensor inputs, historical behavioral data, and real-time signals.

Instead of responding to discrete prompts, the system constructs goal hierarchies. It decomposes high-level intent into sub-tasks, schedules them, resolves interdependencies, and executes across multiple application boundaries.

That requires:

• Cross-app API orchestration

• On-device vector memory persistence

• Local inference acceleration via dedicated NPUs

• Secure enclave processing for private data reasoning

• Latency-optimized decision loops

In short, it’s not just a feature. It’s a systems architecture upgrade.

Why does this matter beyond consumer hardware?

Because what’s happening on-device is a microcosm of enterprise transformation.

Reactive AI reduces execution time per task. Agentic AI reduces coordination complexity across systems.

In most organizations, the bottleneck isn’t computation. It’s orchestration overhead which is humans manually stitching together CRM updates, ERP workflows, analytics dashboards, messaging threads, and compliance checks.

Thus, each handoff introduces latency, context loss, and error propagation.

Agentic systems collapse that fragmentation layer.

That’s the enterprise analog of a phone rebooking your delayed flight and updating your calendar without being asked.

The strategic implication for you is clear.

Thus, the competitive edge will not come from “having AI.” Model access is commoditizing.

The differentiation layer will be orchestration intelligence like how autonomously your systems can execute multi-step objectives across fragmented infrastructure.

Why Perplexity Computer Signals the Next Phase of Agentic AI

With Perplexity AI launching Perplexity Computer, the conversation around agentic AI shifts from “smarter assistants” to “autonomous execution infrastructure.”

Unlike single-model systems such as ChatGPT or Claude, Perplexity Computer coordinates 19 distinct models in parallel.

The architectural philosophy mirrors distributed systems design like specialized nodes optimized for distinct workloads, governed by a central reasoning engine.

At the core sits Opus 4.6 for high-level reasoning, and research-heavy subtasks route to Gemini, long-context retrieval leans on ChatGPT 5.2, and lightweight tasks leverage Grok.

Media generation is offloaded to purpose-built models.

Thus, the system dynamically decomposes a user-defined objective into subtasks, spawns sub-agents, assigns them to the most suitable model, and executes asynchronously inside isolated computing environments.

That isolation is critical. Each workflow instance has its own file system, browser session, tool integrations, and API access. This is not stateless prompt-response logic. It’s persistent, stateful execution with environmental memory.

The real distinction isn’t “multi-model.” It’s workflow autonomy.

Technically, this introduces several structural shifts:

• Model-agnostic orchestration (swappable inference engines)

• Asynchronous task scheduling

• Persistent contextual memory across subtasks

• Sub-agent spawning and lifecycle management

• Tool-native execution (browser, CLI, file system)

In distributed computing terms, it resembles a lightweight autonomous operations layer sitting above heterogeneous AI infrastructure.

Strategically, this matters.

The industry assumption has been that models will commoditize.

But, Perplexity is betting the opposite: specialization will accelerate. Instead of one universal model dominating, the value layer will move to orchestration, coordinating best-in-class components dynamically.

Now zoom in on what this means for you.

If agentic systems can autonomously execute cross-model workflows using the same software interfaces your employees use like browser sessions, file systems, APIs, the constraint shifts from model quality to governance architecture.

Because the leverage here is obvious: subscription-tier “digital employees” running parallel, isolated workflows at machine speed. But so is the risk if autonomy exceeds oversight.

So, the competitive question is no longer “Which model are you using?” It’s “Do you control the orchestration layer?”

If platforms like Perplexity Computer prove viable, the advantage will accrue to organizations that design around autonomous workflow execution, not those that simply integrate chat interfaces into existing processes.

Basically, you don’t need 19 models, but you do need to decide whether your systems are ready for AI that doesn’t just respond, it operates.

Anthropic raises the bar on responsibility, not just capability

As AI systems evolve from chat-based assistants into agentic systems that can browse, write code, execute tools, and complete multi-step workflows autonomously, the responsibility framework around them has to evolve too.

That’s the deeper significance behind Anthropic updating its Responsible Scaling Policy (RSP).

Agentic AI changes the risk equation.

A traditional language model responds to prompts. But, an agentic system can plan, take actions across software environments, call APIs, persist memory, and adapt its strategy over time.

That means risk is no longer limited to “bad outputs.” It now includes unintended actions, workflow-level errors, misuse of tools, and compounding effects across systems.

Anthropic’s AI Safety Levels (ASLs) are built on a conditional logic model: if a system crosses a defined capability threshold, then stricter safeguards must activate.

You can think of it as automated risk escalation tied directly to model capability. As autonomy increases, so do security controls, monitoring requirements, and deployment constraints.

Technically, this is a feedback control system layered on top of capability scaling.

But here’s the complication: capability thresholds are rarely clean or binary. A model might demonstrate partial biological reasoning skills without clearly enabling harm.

Similarly, an agentic system might show early long-horizon planning ability without being fully autonomous. These “gray zones” make it hard to prove whether a system has definitively crossed a risk boundary.

That ambiguity is exactly why structured governance matters.

The updated RSP introduces clearer separation between unilateral safeguards (what a company can enforce on its own) and broader ecosystem-level recommendations (what likely requires industry or government coordination).

It also emphasizes recurring Risk Reports, external expert review, and transparent safety roadmaps. In simple terms: document the risks, publish the mitigation plan, and allow third parties to scrutinize it.

Now let’s translate this for you:

If you’re adopting agentic AI, especially systems that can execute tasks, access internal data, or operate for extended periods without supervision, your responsibility increases alongside capability.

Agentic AI introduces operational leverage, but also operational exposure.

Thus, the key insight is simple: don’t scale autonomy without scaling oversight. Build internal “capability tiers” tied to escalating safeguards. Treat autonomy as something you graduate into, not something you switch on.

Thus, as AI systems become more agentic, governance becomes infrastructure, not policy paperwork. If your AI can act independently, you need equally structured systems to monitor, evaluate, and constrain that independence.

In the agentic era, safety is not about limiting innovation. It’s about ensuring that increased autonomy doesn’t outpace your ability to control it.

If you operationalize structured agent interfaces now, you position your systems to capture agent-mediated demand with higher reliability and lower execution risk.

Your opinion matters!

Hope you loved reading our piece of newsletter as much as we had fun writing it. 

Share your experience and feedback with us below ‘cause we take your critique very critically. 

Thank you for reading

-Shen & Towards AGI team