Nvidia's Q4 Hyperscaler Capex Print Just Reset the Top of the Range

The narrative around Nvidia's revenue trajectory has been bipolar through the trailing 18 months. The bull case argued that the AI capex cycle had multi-year runway, with hyperscaler capex doubling from the 2023 base by 2026. The bear case argued that the AI capex would normalise once the early infrastructure build-out completed, with capex growth decelerating to mid-single digits by 2026. Both cases had supporting data points and the share price oscillated in a wide range as the consensus tried to triangulate.

The Q1-Q3 2025 hyperscaler earnings cycle produced mixed signals. Microsoft and Amazon both raised capex guidance modestly. Alphabet held guidance. Meta accelerated guidance. The aggregate capex pool for calendar 2025 sat at approximately $295 billion, broadly in line with the high end of the consensus model. The data was constructive but not decisive.

The Q4 2025 earnings cycle was the decisive print. Microsoft's calendar 2026 capex guidance of $115-120 billion. Amazon AWS capital expenditure of $130-140 billion. Alphabet's $80-85 billion. Meta's $60-65 billion. Combined $385-410 billion. The aggregate is materially above what the consensus had been modelling and the trajectory points to continued growth in calendar 2027 even at a moderating rate. The floor reset is real.

The second-order signal is that the qualitative commentary across all four hyperscalers emphasised AI infrastructure demand as the dominant driver of capex growth. Microsoft specifically called out Azure AI capacity as the gating factor for revenue growth. Amazon highlighted Bedrock and the broader AI service stack. Alphabet emphasised TPU and GPU capacity for both internal use and Cloud customers. Meta described the Llama infrastructure as multi-year. The qualitative consistency across all four is notable; this is not a single-customer story but a sector-wide capex pool.

The forward revenue trajectory needs to be reconstructed against the hyperscaler capex floor reset. The data centre revenue line has been growing at a 100%+ rate through the trailing twelve months, anchored on the H100 and H200 deployment cycle. The Blackwell architecture launch through 2025-2026 sustains the capacity demand at premium ASPs, with the GB200 system pricing at meaningful premiums to the comparable H100 system on a per-unit-compute basis. The product cycle is supportive of continued ASP expansion alongside unit volume growth.

The sovereign AI demand layer adds a separate revenue stream that the consensus has been slow to incorporate. Saudi Arabia, the UAE, India, Japan, France, and several other governments have committed to multi-billion-dollar AI infrastructure investments through 2026-2028. The cumulative sovereign AI capex pool sits at approximately $80-120 billion through 2027, with Nvidia capturing the majority of the GPU revenue from those programmes. The sovereign demand is incremental to the hyperscaler capex pool and supports continued growth even as the hyperscaler trajectory normalises.

The enterprise AI revenue line is the third leg of the demand stack. Enterprise customers across financial services, healthcare, manufacturing, and government have begun deploying AI infrastructure at scale. The Mellanox networking integration, the Spectrum-X Ethernet platform, and the broader software stack (CUDA, NIM, Omniverse) produce a software-attached revenue model that compounds the hardware revenue. The enterprise contribution is approximately 8-12% of total revenue and growing at high-double-digit rates.

The Q4 2025 hyperscaler capex print is the floor signal that supports the entire revenue trajectory. With the hyperscaler capex pool at $385 billion and Nvidia capturing approximately 70% of the GPU revenue from that pool, the data centre AI revenue line should print $250-280 billion in fiscal 2027. Add the sovereign AI contribution of $20-30 billion, the enterprise contribution of $25-35 billion, and the gaming and professional visualisation lines of $15-20 billion. The aggregate fiscal 2027 revenue trajectory of $290-360 billion is well above the consensus $260-275 billion.

AMD has been the primary competitor at the discrete data centre GPU tier. The MI300X and the MI325X have produced incremental design wins at hyperscalers and at sovereign customers, with the MI325X cumulative bookings sitting in the $5-7 billion range. The competitive position is real but the magnitude is modest against Nvidia's $216 billion revenue base. AMD's market share at the AI training tier remains in the low-single digits; the inference tier has more competitive opportunity but Nvidia's CUDA software moat creates meaningful switching costs.

The internal silicon programmes at the hyperscalers are the second competitive vector. Google's TPU v6 and v7 deployments. Amazon's Trainium and Inferentia chips. Microsoft's Maia silicon. Meta's MTIA accelerators. Each programme represents an attempt to reduce dependence on Nvidia GPUs at the inference layer. The cumulative custom silicon capex sits at approximately $25-35 billion annually across the four hyperscalers. The custom silicon is real but each hyperscaler continues to commit increasing capex pools to Nvidia GPUs alongside the custom programmes; the relationship is supplementary rather than competitive.

The broader competitive landscape includes Cerebras, Groq, SambaNova, and the smaller AI chip startups. Each represents a niche competitive threat at specific workload types but none has scaled to a level that meaningfully challenges Nvidia's position. The capital and software moat that Nvidia has built over the trailing decade is deeper than any single competitor can replicate within a 24-month window. The competitive position at the GPU layer is structurally durable.

The Signals Desk read on the competitive set is that Nvidia's market share at the AI training tier will remain above 85% through 2027, with modest erosion at the inference tier offset by the CUDA software stack and the systems-level integration of networking and software. The competitive concern is a multi-year story rather than a near-term catalyst.

The institutional flow data through Q1 2026 shows the bid building rather than fading. Top-twenty institutional holders added net positions through the back half of 2025, and the sovereign wealth fund allocations expanded measurably. The retail flow has been more cautious, with Robinhood and similar retail-focused brokerages showing position trimming. The combination of institutional accumulation against retail caution is the textbook pattern for the next leg of a multi-quarter outperformance window.

The options market positioning has evolved alongside the capex disclosure. The implied volatility on the 1-month options has compressed by approximately 8 percentage points since the Q4 disclosure cycle, indicating that the market views the trajectory as more predictable. The downside skew has steepened modestly, reflecting tail-risk hedging demand, but the broader options structure is consistent with continued upward bias rather than a topping pattern.

The technical setup is supportive. The 50-day moving average sits at $185 with the 200-day at $182. The crossover signal is constructive. The 52-week range of $104 to $212 captures the volatility through the prior year, but the volume profile shows institutional accumulation in the $170-190 zone over the second half of 2025. That zone has been the institutional buying line and the technical breakdown signal has not been triggered.

Free cash flow conversion is the underrated variable in the Nvidia trajectory. The fiscal 2026 FCF print of $96.7 billion against revenue of $216 billion is a 45% conversion ratio, an extraordinary outcome for a hardware-led business at this scale. The conversion reflects the limited working capital intensity of the GPU model, the modest capex base, and the high gross margin profile. Each of those features compounds the cash flow line at a faster rate than the revenue line, which produces the operating leverage benefit at the per-share level.

Apply the FCF trajectory to the share count progression. Nvidia has been buying back stock at an accelerating pace, with the trailing twelve months buyback execution at approximately $40 billion. The combination of the FCF expansion and the share retirement produces FCF per share growth that exceeds the consolidated FCF growth. The per-share trajectory is what matters for the equity return, and the trajectory is favourable through the next 24 months at minimum.

The balance sheet positioning provides additional flexibility. Nvidia held approximately $43 billion of cash and equivalents at year-end fiscal 2026 against effectively zero debt. The net cash position represents roughly 0.85% of market cap, smaller than it sounds in proportional terms but meaningful in absolute terms. The cash provides optionality for both incremental buyback acceleration and strategic acquisition activity in the AI infrastructure or AI software stack. The capital allocation discipline through the trailing eighteen months has been evident.

The FCF conversion at this scale also creates a structural support for the multiple. At $96 billion of FCF on a $5 trillion market cap, the FCF yield sits at approximately 1.9%. The yield is modest but the growth rate is the variable that supports the multiple. With FCF on track to compound at 30-40% annually over the next two years, the FCF yield expands rapidly even at flat share prices. The mathematics of the compounding are what holds the multiple in place against incremental volatility.

The sovereign AI demand layer deserves a closer look because it is one of the most under-modelled components of the Nvidia revenue trajectory. The Saudi PIF-funded Humain programme has committed to multi-billion-dollar GPU procurement through 2027. The UAE G42 partnership has scaled to a meaningful Nvidia revenue contributor. The Indian programme under the IndiaAI mission has committed to $1.5-2.0 billion of GPU procurement. The Japanese METI programme has allocated similar capital. France's national AI capacity programme has begun at a smaller scale but with explicit multi-year capex commitment. The cumulative sovereign AI commitment pool sits at $80-120 billion through 2027.

The sovereign demand is structurally different from the hyperscaler demand. The contracts are typically multi-year, with capacity scheduling that smooths the revenue recognition. The customer concentration risk is lower because the sovereign programmes are policy-driven rather than commercial-cycle-driven. The pricing flexibility is higher because the sovereign programmes have explicit cost-of-capacity targets rather than open-market pricing dynamics. Each feature of the sovereign demand profile reduces the volatility of the underlying revenue line.

The geopolitical complexity is the meaningful tail risk. US export controls on AI accelerators have constrained the Chinese market opportunity. Any tightening of the export envelope to other jurisdictions would compress the sovereign demand profile. The current US administration has been broadly supportive of the sovereign AI partnerships in friendly markets but the policy continuity over a multi-year horizon is not guaranteed. We are conservatively modelling the sovereign AI revenue contribution at the lower end of the range to capture the policy uncertainty.

The other piece of the sovereign demand worth flagging is the multi-tenant data centre operator layer. Stargate, the partnership between OpenAI, Oracle, SoftBank, and others, has committed to multi-hundred-billion-dollar AI infrastructure investment over the coming years. Coreweave, Lambda Labs, Nebius, and the broader GPU-as-a-service operator set have all expanded capacity commitments measurably. The combined demand from this layer adds another $30-50 billion of incremental Nvidia revenue through 2027, which is an additional incremental demand beyond the hyperscaler and sovereign pools.