The AI software vs hardware debate is no longer academic. After two years of hardware-led gains, the 2026 setup looks like the start of a genuine capital rotation toward software monetization.
Hyperscaler capex is still rising, but the marginal dollar is asking a harder question. Where does the revenue come from once the GPUs are installed?
This piece walks the two phases, the signals that mark the handoff, and how to construct an AI sleeve that owns both sides.
Phase 1: Hardware Boom (NVDA, AVGO, AMD) 2024 To 2025
Phase 1 was a buildout cycle. Every hyperscaler raced to lock in compute, and chip suppliers absorbed almost the entire profit pool.
Nvidia captured the lion's share
Nvidia (NVDA) exited fiscal 2026 with record quarterly revenue of US$68.1 billion, up 73% year over year per its investor relations filings. Data center remained the engine.
Blackwell shipments anchored hyperscaler cluster builds through 2025, and CUDA kept switching costs punishingly high for rivals.
Broadcom and AMD took the rest
Broadcom (AVGO) rode custom accelerator wins with Google and Meta, while AMD finally gained meaningful MI300 share at Microsoft and Oracle.
The trade was simple. Buy the suppliers, ride the capex curve, and ignore the unresolved question of end-customer revenue.
Phase 2: Software Monetization (PLTR, NOW, CRM) Begins
Phase 2 looks different. The installed base of GPUs is now large enough that monetization, not deployment, becomes the bottleneck.
Palantir is the breakout case
Palantir (PLTR) reported Q1 2026 revenue growth of 85% year over year, with US commercial up 104%, per its Q1 2026 shareholder letter. Full-year guidance was raised to roughly 71% growth.
Our deeper read of those numbers lives in our Palantir Q1 2026 review. The federal moat is wider than skeptics modelled.
ServiceNow and Salesforce monetize through workflows
ServiceNow (NOW) is layering AI agents on top of an existing enterprise workflow base of more than 8,000 customers, with measurable seat-level pricing uplift.
Salesforce (CRM) is selling Agentforce into the same installed base, attaching AI revenue to contracts that already exist. Distribution is the moat.
Rotation Signals: Revenue Per Token, Agent Adoption Curves
Calling a rotation early is the easy mistake. The hard work is identifying which signals confirm the handoff is real.
Revenue per token is the cleanest signal
Revenue per token measures how much value the application layer extracts from each unit of inference. When this ratio rises across hyperscaler earnings calls, software is monetizing faster than compute.
The inverse signal is hyperscaler capex growth decelerating while AI revenue accelerates. That gap closes in software's favor.
Agent adoption curves are the second signal
Watch named-account agent rollouts at PLTR, NOW, and CRM. Each agent that replaces a workflow rather than augmenting one is a step-change in pricing power.
For broader context, see our 5 AI stocks beyond Nvidia piece on names with similar adoption tailwinds.
Investors who own only chip suppliers are betting Phase 2 never arrives. A small allocation to PLTR hedges that bet without abandoning the hardware sleeve.
Software Winners With Defensible Distribution Moats
Not every software vendor will benefit. The winners are platforms that already own customer workflows.
Distribution beats model quality
Model quality converges fast. Distribution does not. A platform sitting inside Fortune 500 procurement, security review, and integration stacks can ship a mediocre agent and still win the contract.
That is the PLTR, NOW, and CRM advantage. They sell into installed bases that took a decade to build.
Pricing power shows up in net revenue retention
Net revenue retention above 120% signals customers expanding spend without churn. PLTR US commercial is running well above that threshold today.
If a software name is selling AI without lifting NRR, the moat is thinner than the multiple implies. That is the screen.
Balanced AI Sleeve Construction For A Long-Term Portfolio
A long-term AI sleeve should own both layers and rebalance as signals shift. Single-layer concentration is the avoidable risk.
One workable structure is 50% hardware (NVDA, AVGO, AMD), 35% software (PLTR, NOW, CRM), and 15% cash for tactical adds on drawdowns.
Rebalance quarterly. Trim hardware on capex deceleration, add software on rising revenue-per-token. The structure forces discipline when narratives swing.
Conclusion
The AI software vs hardware rotation is not a binary call. It is a gradual rebalancing as monetization signals catch up to the buildout.
Hardware will remain a core holding because compute demand is not slowing. Software is the layer that monetizes the compute already installed, and that is where the next leg of returns likely lives.
You can build a balanced AI sleeve on Gotrade with fractional shares from US$1.
Start by sizing a small position in PLTR alongside your existing chip exposure, then rebalance as the rotation plays out.
FAQ
Is the AI hardware trade over?
No. Hardware demand stays strong while hyperscaler capex grows, but the marginal dollar of return is shifting toward software monetization layers.
Why is Palantir the lead software name in this rotation?
PLTR posted 85% revenue growth in Q1 2026 with US commercial up 104%, the cleanest evidence that enterprise AI is monetizing through entrenched distribution.
How do I track the rotation in real time?
Watch revenue per token in hyperscaler disclosures, agent adoption rates at PLTR, NOW, and CRM, and net revenue retention above 120% for software names.
What allocation makes sense for a long-term AI sleeve?
A practical starting mix is 50% hardware, 35% software, and 15% cash for tactical rebalancing, sized at 5% to 15% of your equity portfolio.
