Revisiting Our Nvidia Thesis: The Bull Case Just Got Louder

The original five trillion narrative made several specific predictions. First, that AI capex by the hyperscaler cohort would continue growing at 35-50 percent annually through CY2026. The actual hyperscaler capex growth has been approximately 45 percent year-over-year on aggregate. Within the predicted range. Right.

Second, the original thesis predicted that Nvidia's data centre revenue would scale to approximately $180-200 billion by FY26. The actual data centre revenue for FY26 reached approximately $185 billion. Within the predicted range. Right.

Third, the original thesis predicted that the operating margin would expand to 58-62 percent. The actual FY26 operating margin was 60 percent. Within the predicted range. Right.

Fourth, the original thesis predicted that free cash flow conversion would exceed 75 percent of net income. The actual FY26 free cash flow conversion was approximately 80 percent. Slightly above the predicted range. Right and modestly better.

Where the original thesis underestimated: the speed of inference-side capex acceleration. The original framework had inference capex contributing approximately 25 percent of total AI infrastructure spend by CY2026. The actual share has trended toward 35-40 percent as customer-facing AI applications have scaled faster than the model anticipated. The inference share matters because the unit economics on inference compute are slightly different from training compute, and the addressable market is larger.

The FY26 outperformance was driven principally by inference capex acceleration. The pattern argues for a multi-year extension of the capex cycle rather than the cyclical peak that some consensus voices have been suggesting.

This week's news flow reinforced the Nvidia thesis on multiple fronts. The headline 'Wall Street's Biggest AI Bull Just Got Louder' captured the most prominent bullish update from a top-tier sell-side analyst raising the price target meaningfully. Multiple smaller research updates followed in the same direction. The price action on the stock during the week was constructive despite a broader equity market under modest pressure.

The Signals Desk read on the news flow pattern is that the consensus has been progressively catching up to the operational data rather than running ahead of it. The original thesis identified a gap between the operational trajectory and consensus expectations. That gap has been closing but has not yet fully closed. The remaining gap is the source of the continued upside potential.

The specific consensus revisions over the past quarter have moved Wall Street's FY27 EPS estimate up by approximately 12 percent. The Signals Desk forward EPS forecast is still 8-10 percent above the updated consensus. The gap is narrower than at the time of the original publication but remains meaningful.

The single biggest update to the Signals Desk view since the original thesis is the shape of the inference capex wave. Inference compute is the demand for AI processing power that supports running already-trained models at production scale. Every consumer-facing AI application, every enterprise AI deployment, and every embedded AI feature consumes inference compute. The demand scales with active usage rather than with training cycles.

The data over the past nine months shows that inference compute demand is growing at 80-100 percent year-over-year. The growth rate is higher than training compute demand because the deployment of trained models is happening faster than the training of new models. Every major hyperscaler has shifted capex allocation toward inference-oriented hardware in their forward guidance.

Nvidia's inference-oriented chip portfolio includes the H200, the L40S for inference workloads, the Grace-Blackwell platform, and the recently announced Rubin family. The competitive position in inference is slightly less unassailable than in training, with AMD's MI300 and MI400 platforms achieving some traction in inference workloads and the hyperscaler-designed custom silicon (Trainium, Inferentia, Maia, MTIA) capturing meaningful internal workloads.

The market dynamics in inference are therefore: total demand growing at 80-100 percent annually, Nvidia share approximately 75-85 percent (compared to 90 percent plus in training), and competitive intensity gradually increasing as alternative silicon scales. The net effect for Nvidia is still extraordinary growth, but with slightly more competitive pressure on price per chip than in the training-dominated phase.

The Signals Desk model now assumes Nvidia data centre revenue scales to approximately $280-320 billion by FY28. The FY26 actual of $185 billion is the launching point. The growth path implies continued 40-60 percent revenue growth in FY27 and 25-40 percent in FY28.

The updated Signals Desk model produces the following revised framework. FY27 (year ending January 2027) revenue forecast: $290-310 billion. FY27 operating margin: 58-61 percent. FY27 operating income: $170-190 billion. FY27 net income: $150-170 billion. FY27 free cash flow: $130-150 billion. FY27 EPS: approximately $6.20-7.00 (post the recent stock split). FY28 EPS: approximately $8.00-9.50 if the inference capex wave continues at the current trajectory.

At the current $233 share price (post-split adjusted), the forward P/E on FY27 EPS sits at 33-37 times. On FY28 EPS, the forward P/E compresses to 24-29 times. Neither multiple is excessive for the implied growth rate.

Applying a 35-40 times multiple to FY27 EPS produces a price target range of $215-280 within 12 months. Applying a 28-32 times multiple to FY28 EPS produces a target range of $225-305 within 24 months. The Signals Desk updated target is $245-275 over a 24-month horizon.

The principal upside risk to the forecast is faster-than-expected inference capex acceleration combined with sovereign AI demand expansion. Sovereign demand, which is governments and state-backed enterprises building national AI infrastructure, has been accelerating since CY2025. The realised contribution to FY27 revenue could exceed $30 billion if the announced sovereign initiatives translate to procurement.

The principal downside risk to the forecast is hyperscaler capex pause if AI revenue monetisation fails to keep pace with infrastructure spend. The Signals Desk view on this risk is that the cycle is far enough along that a pause is unlikely before late CY2027 at the earliest.

Since the original five trillion piece was published, the sovereign AI demand layer has become a meaningful new contributor to the data centre revenue mix. Multiple national governments have announced AI infrastructure initiatives backed by direct procurement commitments to Nvidia.

The specific announced commitments include programs in Saudi Arabia, UAE, Singapore, Japan, France, UK, India, Korea, and Australia. The cumulative announced procurement value exceeds $80 billion across a multi-year horizon. The realised revenue contribution depends on procurement timing, export licence approvals, and political stability.

The sovereign AI demand has two structural advantages for Nvidia. First, sovereign customers tend to pay premium pricing because they prioritise security of supply over absolute lowest cost. Second, sovereign customers tend to commit to multi-year procurement that smooths the cyclical demand profile compared to the hyperscaler cohort.

The Signals Desk model now incorporates sovereign demand as a distinct revenue segment growing at 40-60 percent annually from a base of approximately $15 billion in FY26 to potentially $35-50 billion by FY28. The sovereign segment provides a meaningful diversification of the customer concentration risk that was a focal point of bear arguments in 2024.

The original thesis predicted that Nvidia would expand the capital return program in line with the cash flow scale. The actual program has expanded materially. The FY26 capital return included approximately $42 billion of buybacks and a modest dividend. The dividend is small as a percentage but is the first formal recurring capital return from the franchise. The buyback authorisation expansion to over $50 billion confirmed in the FY26 print exceeds the original thesis expectations.

The share count reduction profile is now meaningful. Nvidia has reduced share count by approximately 3-4 percent annually through the recent buyback execution. That share count reduction supports EPS compounding at a rate above the operating income growth rate.

The balance sheet position remains strong. Net cash of approximately $40 billion at the end of FY26. The cash position grows by the difference between free cash flow generation and capital return, plus the impact of any strategic M&A or investment activity. The Signals Desk view is that the cash position will continue to build through FY28 as free cash flow growth outpaces the buyback expansion.

The optionality on the cash position includes selective M&A in the software, infrastructure, and emerging compute platform spaces. Recent acquisitions have been small (Run.ai, OmniML, several others) and focused on capability expansion. The Signals Desk does not anticipate large transformational M&A but expects continued bolt-on activity.

Three risks have evolved meaningfully since the original thesis was published. First, the customer concentration risk has been clarified. The top five hyperscaler customers continue to account for approximately 45-50 percent of data centre revenue. The concentration has been stable rather than increasing. The diversification across sovereign AI, enterprise AI, and the broader cloud-services tier has been faster than the original thesis assumed.

Second, the competitive risk from custom silicon has been increasing. Amazon's Trainium 3, Google's TPU v7, Microsoft's Maia 200, and Meta's MTIA 3 have all been deployed at scale. Each captures a portion of internal training and inference workloads. The Signals Desk view is that custom silicon will capture approximately 25-30 percent of hyperscaler internal compute by FY28, up from approximately 15-20 percent in FY26. The growth in custom silicon does not collapse the total addressable market for Nvidia, but it does compress the percentage of incremental hyperscaler capex going to merchant silicon.

Third, the geopolitical risk has shifted. Export controls on advanced AI chips have continued to evolve. Nvidia has lost the China market for the most advanced chips but has retained the China market for downgraded variants. The China revenue contribution remains material but has been depressed compared to the pre-export-control baseline. The Signals Desk model assumes China revenue stays in the $10-15 billion range annually through FY28.

The net effect of the three risks is that the bull case has been reinforced more than the bear case. The customer concentration is stable. The custom silicon competition is real but not catastrophic. The geopolitical situation has stabilised at a manageable level.