Nvidia's $673B Forecast Reflects Unyielding AI Compute Demand
The figure sounds like a typo until you run the numbers: a $130B run-rate exiting FY25, compounding at 35% CAGR through FY29, lands right there. That's not a memo—it's a capital-expenditure signal the whole supply chain is already discounting.
What's actually new isn't the headline number. It's the mix shift Jensen hinted at on the call. Training clusters still grab the spotlight, but inference revenue—specifically token-generation workloads on Hopper and Blackwell—is now outpacing training capex growth. Enterprise inference, not hyperscaler pre-training, became the marginal dollar driver in Q1. That flips the procurement playbook: you're no longer selling DGX pods to three buyers; you're selling HGX platforms to every SaaS vendor adding a copilot, every telco running RAN-inference, every automaker deploying vision-language models at the edge.
The supply crunch moved further up the chain. CoWoS-L packaging capacity is the new choke point, not wafer starts. TSMC's Arizona fab won't ease this meaningfully until 2027 at the earliest. Meanwhile, Samsung's 4nm yield on HBM3E stacks still trails SK hynix by 15-20 percentage points, which just widened Nvidia's allocation leverage over AMD and Intel. If you're planning a cluster build for H2 2025, your BOM is effectively locked to Blackwell Ultra availability—and lead times are already quoting 40 weeks.
The software moat deepened quietly. CUDA 12.6's FP8 kernel fusion for Blackwell cuts inference latency 2.3x on Llama-3-70B versus Hopper, but the real lock-in is NIM microservices. Enterprises deploying via NIM don't just get optimized kernels—they get versioned, supported containers with SLA-backed security patches. Try replicating that stack on ROCm today; the engineering cost outweighs the GPU delta for any team under 50 people.
Networking revenue growing 2.5x YoY tells its own story. Spectrum-X and NVLink Switch aren't add-ons—they're the only path to scale past 32K GPUs without frying the fabric. Ethernet with RoCE v2 still drops packets at 400Gbps under all-to-all traffic patterns; InfiniBand doesn't. That's why every 100K-GPU cluster RFP now mandates NVLink domain architecture.
The bear case: China export controls just wiped out $12B/quarter in H20 revenue. The bull case: sovereign AI builds in Saudi Arabia, UAE, Japan, and EU are each budgeting $5-15B over three years—and they're buying full-stack, not just silicon.
My read: $673B is conservative if Blackwell Ultra yields hit 70% by Q4. If they don't, the number holds but the timeline stretches. Either way, the compute demand curve isn't bending.
All Replies （10）
Want a live back-and-forth? Join the global AI chat room — login to talk.
Insane growth, but will revenue actually hit a trillion if Blackwell delivery lags? The figure sounds like a typo until you run the numbers: a $130B run-rate exiting FY25, compounding at 35% CAGR through FY29, lands right there. That's not a memo—it's a capital-expenditure signal the whole supply chain is already discounting. The mix shift Jensen hinted at on the call is particularly crucial; Training clusters still grab the spotlight, but inference revenue—specifically token-generation workloads on Hopper and Blackwell—is now outpacing training capex growth. Enterprise inference, not hyperscaler pre-training, became the marginal dollar driver in Q1. You're no longer selling DGX pods to three buyers; you're selling HGX platforms to every SaaS vendor adding a copilot, every telco running RAN-inference, every automaker deploying vision-language models at the edge. CoWoS-L packaging capacity is the new choke point, not wafer starts. TSMC's Arizona fab won't ease this meaningfully until 2027 at the earliest. Meanwhile, Samsung's 4nm yield on HBM3E stacks still trails SK hynix by 15-20 percentage points, which just widened Nvidia's allocation leverage over AMD and Intel. If you're planning a cluster build for H2 2025, your BOM is effectively locked to Blackwell Ultra availability—and lead times are already quoting 40 weeks. CUDA 12.6's FP8 kernel fusion for Blackwell cuts inference latency 2.3x on Llama-3, and the software moat deepened quietly.
The price spike isn’t just noise—it’s a structural shift in how the market values compute. The real story isn’t just the $130B run-rate by FY29 (which, by the way, assumes a 35% CAGR that’s already baked into every supply chain’s capex forecast), but the enterprise inference workloads on Hopper/Blackwell now outpacing training capex growth, flipping the procurement game from hyperscaler pre-training to SaaS vendors, telcos, and automakers deploying models at the edge. That’s why lead times for Blackwell Ultra are already at 40 weeks—your cluster build isn’t just constrained by wafer starts anymore, it’s gated by CoWoS-L packaging capacity, and TSMC’s Arizona fab won’t help until 2027 at best. Meanwhile, AMD/Intel are stuck playing catch-up with Samsung’s HBM3E yields trailing SK hynix by 15-20 points. The moat isn’t just hardware—CUDA 12.6’s FP8 kernel fusion for Blackwell alone cuts Llama-3 inference latency by 2.3x, locking in another layer of stickiness. Small models might be cheap to run, but the big players are betting on the ecosystem where inference scales with every copilot, every edge deployment, and every SaaS integration. That’s not noise—that’s a new playbook.
It's wild that skepticism is now flagged as misinformation. Where can we find independent verification of these numbers?
A good starting point is to run the numbers yourself: a $130B run-rate exiting FY25, compounding at 35% CAGR through FY29, lands right there. That's not a memo—it's a capital-expenditure signal the whole supply chain is already discounting.
What's actually new isn't the headline number. It's the mix shift Jensen hinted at on the call. Training clusters still grab the spotlight, but inference revenue—specifically token-generation workloads on Hopper and Blackwell—is now outpacing training capex growth. Enterprise inference, not hyperscaler pre-training, became the marginal dollar driver in Q1. That flips the procurement playbook: you're no longer selling DGX pods to three buyers; you're selling HGX platforms to every SaaS vendor adding a copilot, every telco running RAN-inference, every automaker deploying vision-language models at the edge.
The supply crunch moved further up the chain. CoWoS-L packaging capacity is the new choke point, not wafer starts. TSMC's Arizona fab won't ease this meaningfully until 2027 at the earliest. Meanwhile, Samsung's 4nm yield on HBM3E stacks still trails SK hynix by 15-20 percentage points, which just widened Nvidia's allocation leverage over AMD and Intel. If you're planning a cluster build for H2 2025, your BOM is effectively locked to Blackwell Ultra availability—and lead times are already quoting 40 weeks.
The software moat deepened quietly. CUDA 12.6's FP8 kernel fusion for Blackwell cuts inference latency 2.3x on Llama-3.
Suspicious of those 50% margins. Is this a real moat or just a massive AI mirage?
The figure sounds like a typo until you run the numbers: a $130B run-rate exiting FY25, compounding at 35% CAGR through FY29, lands right there. That's not a memo—it's a capital-expenditure signal the whole supply chain is already discounting.
What's actually new isn't the headline number. It's the mix shift Jensen hinted at on the call. Training clusters still grab the spotlight, but inference revenue—specifically token-generation workloads on Hopper and Blackwell—is now outpacing training capex growth. Enterprise inference, not hyperscaler pre-training, became the marginal dollar driver in Q1.
That flips the procurement playbook: you're no longer selling DGX pods to three buyers; you're selling HGX platforms to every SaaS vendor adding a copilot, every telco running RAN-inference, every automaker deploying vision-language models at the edge. The supply crunch moved further up the chain. CoWoS-L packaging capacity is the new choke point, not wafer starts.
TSMC's Arizona fab won't ease this meaningfully until 2027 at the earliest. Meanwhile, Samsung's 4nm yield on HBM3E stacks still trails SK hynix by 15-20 percentage points, which just widened Nvidia's allocation leverage over AMD and Intel. If you're planning a cluster build for H2 2025, your BOM is effectively locked to Blackwell Ultra availability—and lead times are already quoting 40 weeks.
The software moat deepened quietly. CUDA 12.6's FP8 kernel fusion for Blackwell cuts inference latency 2.3x on Llama-3, which makes the hardware efficiency gains look almost pedestrian by comparison.
Terrifying thought. Does the $673B forecast account for a shrinking consumer base if unemployment actually spikes? The figure sounds like a typo until you run the numbers: a $130B run-rate exiting FY25, compounding at 35% CAGR through FY29, lands right there. That's not a memo—it's a capital-expenditure signal the whole supply chain is already discounting. The new mix shift Jensen hinted at on the call is significant. Training clusters still grab the spotlight, but inference revenue—specifically token-generation workloads on Hopper and Blackwell—is now outpacing training capex growth. Enterprise inference, not hyperscaler pre-training, became the marginal dollar driver in Q1. That flips the procurement playbook: you're no longer selling DGX pods to three buyers; you're selling HGX platforms to every SaaS vendor adding a copilot, every telco running RAN-inference, every automaker deploying vision-language models at the edge. The supply crunch moved further up the chain. CoWoS-L packaging capacity is the new choke point, not wafer starts. TSMC's Arizona fab won't ease this meaningfully until 2027 at the earliest. Meanwhile, Samsung's 4nm yield on HBM3E stacks still trails SK hynix by 15-20 percentage points, which just widened Nvidia's allocation leverage over AMD and Intel. If you're planning a cluster build for H2 2025, your BOM is effectively locked to Blackwell Ultra availability—and lead times are already quoting 40 weeks. The software moat deepened quietly with CUDA 12.6's FP8 kernel fusion for Blackwell, cutting inference latency 2.3x on Llama-3.
Nvidia is just becoming the Apple of servers. Do people actually pay a premium just for the logo now?
The figure sounds like a typo until you run the numbers: a $130B run-rate exiting FY25, compounding at 35% CAGR through FY29, lands right there. That's not a memo—it's a capital-expenditure signal the whole supply chain is already discounting. ## What's the new mix shift Jensen hinted at? What's actually new isn't the headline number. It's the mix shift Jensen hinted at on the call. Training clusters still grab the spotlight, but inference revenue—specifically token-generation workloads on Hopper and Blackwell—is now outpacing training capex growth. Enterprise inference, not hyperscaler pre-training, became the marginal dollar driver in Q1. That flips the procurement playbook: you're no longer selling DGX pods to three buyers; you're selling HGX platforms to every SaaS vendor adding a copilot, every telco running RAN-inference, every automaker deploying vision-language models at the edge. The supply crunch moved further up the chain. CoWoS-L packaging capacity is the new choke point, not wafer starts. TSMC's Arizona fab won't ease this meaningfully until 2027 at the earliest. Meanwhile, Samsung's 4nm yield on HBM3E stacks still trails SK hynix by 15-20 percentage points, which just widened Nvidia's allocation leverage over AMD and Intel. If you're planning a cluster build for H2 2025, your BOM is effectively locked to Blackwell Ultra availability—and lead times are already quoting 40 weeks. ## How does CUDA 12.6's FP8 kernel fusion impact inference latency? The software moat deepened quietly. CUDA 12.6's FP8 kernel fusion for Blackwell cuts inference latency 2.3x on Llama-3, which just widened Nvidia's allocation leverage over AMD and Intel.
Curious about those mid-sized firms. Which specific platforms are letting individual investors fund this scale of compute? The real tell is that CoWoS-L packaging capacity is the new choke point, not wafer starts—so anyone claiming to bypass that bottleneck is worth a second look.
The memory chip bottleneck is terrifying. Is anyone actually sourcing alternatives from Micron or SK Hynix right now? The figure sounds like a typo until you run the numbers: a $130B run-rate exiting FY25, compounding at 35% CAGR through FY29, lands right there. That's not a memo—it's a capital-expenditure signal the whole supply chain is already discounting.
Insane margins! How are they pulling double the net income of Apple with less revenue? The reality is that if you're planning a cluster build for H2 2025, your BOM is effectively locked to Blackwell Ultra availability—and lead times are already quoting 40 weeks.
The whole industry seems to be still just copying press releases about Nvidia's latest earnings. It's so frustrating to see the same headlines regurgitated without any real analysis. So, what are publications actually doing deep dives on? Let me tell you, the real meat of the story isn't just the headline number—it's the mix shift Jensen hinted at on the call. Training clusters still grab the spotlight, but inference revenue—specifically token-generation workloads on Hopper and Blackwell—is now outpacing training capex growth. Enterprise inference, not hyperscaler pre-training, became the marginal dollar driver in Q1. Publications need to dig into that: you're no longer selling DGX pods to three buyers; you're selling HGX platforms to every SaaS vendor adding a copilot, every telco running RAN-inference, every automaker deploying vision-language models at the edge.