The Number Microsoft Wants You to Celebrate and the One It Buried
Microsoft processed over 100 trillion tokens in the quarter ending April 2025. Token usage was up 5x year over year. In the final month of that quarter alone, the company processed a record 50 trillion tokens. Those are the numbers that made headlines and lit up VC Twitter.
Here's the number Microsoft buried in the same earnings call: "The real outperformance in Azure this quarter was in our non-AI business."
That quote does a lot of work. It signals that despite the token milestone, despite the infrastructure build, despite the 5x growth in consumption the AI business itself is not yet the engine. It's the sidecar. Impressive, fast growing, strategically critical and not yet the financial center of gravity.
This briefing is about what 100 trillion tokens actually reveals when you run the math end to end: from Microsoft's infrastructure bet, to the real cost structure for enterprise buyers, to the pricing threat coming from China that most Western analysts are only beginning to take seriously.
---
What Does 5x Token Growth Actually Mean for Microsoft's Business Model?
5x year over year token growth at Microsoft's scale means the company has executed a successful pivot from selling raw compute to operating what analysts at SemiAnalysis call a "token factory" a business model with a fundamentally different ROIC profile than bare metal cloud contracts. The shift matters because token based revenue is more defensible, more recurring, and harder to migrate off than raw VM pricing. But the margins on getting there are brutal.
Microsoft's token volume growth was supported by real infrastructure execution. The company:
- Cut cost per token by more than half in a single quarter
- Improved AI performance by nearly 30% per unit of power on an ISO power basis
- Reduced GPU dock to lead times by nearly 20% across its blended fleet
- Achieved 1.1 million tokens per second on a single rack (Azure ND GB300 v6, running Llama 2 70B verified by third-party firm Signal65)
That last data point is worth pausing on. One rack. 1.1 million tokens per second. The prior record, also set by Microsoft, was 865,000 tokens per second on the NVIDIA GB200 NVL72. These aren't marketing benchmarks Signal65 independently validated the GB300 result under MLPerf Inference v5.1 workload conditions.
So the infrastructure story is unambiguously strong. The business model story is more complicated.
SemiAnalysis notes that Microsoft's OpenAI training clusters scaled from tens of megawatts to gigawatts an infrastructure transformation that involves an estimated ~$80 billion in AI data center investment in the noted year. That capex has to earn its keep. And the revenue range analysts can actually pin to Azure AI right now runs from a low-scenario $327.6 million annual run rate to a high scenario of $4.6 billion a spread so wide it signals how much is still genuinely unknown about the monetization trajectory.
The honest answer is: this is a massive bet on a business model that is working at the volume level and still being proven at the margin level.
---
Who's Actually Winning the Token Volume War and It Might Not Be Who You Think
By raw daily token throughput, ByteDance's Doubao model is running away with the volume crown at over 120 trillion tokens per day, it dwarfs every Western player by a factor that should be causing genuine strategic discomfort in Redmond and Mountain View
To put that in context:
- Google Gemini APIs: ~14 trillion tokens/day (Q4 2025)
- OpenAI: ~8.6 trillion tokens/day
- Microsoft: processed 500+ trillion tokens in all of H1 2025, or roughly 2.7 trillion tokens/day averaged though the April 2025 month peak implies run rate of ~1.67 trillion/day at that point
- ByteDance Doubao: >120 trillion tokens/day
One commenter on the Tunguz LinkedIn thread estimated that Google does approximately 574x Microsoft Foundry's April token volume when all Google services are included a stat that reframes the "Microsoft wins" narrative significantly. That estimate is unverified and should be read as illustrative rather than definitive, but even a fraction of that gap is meaningful.
China's total AI token usage reached 140 trillion tokens per day as of March 2026, per OpenRouter data cited by Chinese state media. Chinese AI models held 61% market share on OpenRouter as of February 2026.
This isn't a future threat. It's a current-state competitive reality.
---
What Does the China Pricing Gap Actually Do to Western Enterprise Economics?
The pricing gap between Chinese and Western AI models is not a minor market inefficiency it is a structural threat to the premium pricing that Western AI vendors need to fund their R&D and infrastructure investments. DeepSeek V3.2 output tokens were priced at $0.42 per million tokens. Anthropic's Claude Opus 4.6 was listed at $75 per million output tokens. That is a 178x price differential at list pricing.
That gap matters because enterprise buyers are rational. When Microsoft canceled the majority of its own internal Claude Code licenses in early 2026 due to unmanageable costs from employee adoption at scale, it was making exactly the same calculation any CFO would make. When Uber consumed its entire 2026 AI budget in four months, the response wasn't to celebrate AI productivity it was to implement usage caps.
The aggregate data confirms this is not isolated. Average monthly enterprise AI spend rose from $63,000 in 2024 to $85,500 in 2025, a 36% increase. The share of companies spending over $100,000 per month more than doubled in the same period. That's budget shock, not healthy adoption.
And this is before agentic AI workflows hit at scale. The Tomasz Tunguz LinkedIn analysis estimated that agents within GitHub Copilot, Visual Studio, Copilot Studio, and Microsoft Fabric together contribute less than 1% of overall Azure AI inference volume. When agents start running at meaningful scale, token consumption per workflow explodes and so does the bill.
Goldman Sachs projects 24x token consumption growth by 2030. If enterprise pricing doesn't compress dramatically, the usage caps and license cancellations seen in 2025-2026 will look like a warning shot.
---
Is the "Efficiency Win" Narrative Actually Holding Up at the Margin Level?
The efficiency gains are real, but they're not fast enough to prevent margin compression at least not yet. Microsoft's CFO Amy Hood explicitly called out GitHub Copilot compute consumption as a material driver of gross margin decline in the April 2026 earnings call: "Gross margin percentage decreased year over year, driven by continued AI investment and increased GitHub Copilot usage, partially offset by ongoing efficiency gains in Azure."
That's the CFO of a company that just reported a 30% AI performance per watt improvement, a >50% cost per token reduction, and 5x token growth and she's still citing AI as a margin headwind. The efficiency gains are being outpaced by adoption volume. This is Jevons Paradox at cloud scale: make something cheaper and people use more of it, faster than the cost reduction saves you.
The OpenAI unit economics picture from the GPT-5 era makes the same point from a different angle. According to Exponential View's analysis, OpenAI was likely profitable on compute costs alone during the GPT-5 period revenue exceeded direct compute expense. But after accounting for staff, sales and marketing, administrative overhead, and Microsoft revenue sharing arrangements, margins were thin to negative. More damning: R&D spending in the four months before GPT-5's release was likely larger than total gross profits earned during the entire GPT-5 and GPT-5.2 tenure.
This is the real unit economics picture of frontier AI: compute margins are improving, but the full-stack cost structure training, R&D, distribution, partnerships is a different animal. Vendors who survive the next phase will be the ones who can widen the gap between inference revenue and total operating cost, not just inference revenue and compute cost.
Microsoft's June 2026 shift to token based pricing for GitHub Copilot converting fixed $15–100/month plans into monthly usage credits that draw against actual API consumption from Anthropic, OpenAI, Microsoft, and others is the clearest signal yet that the company understands the math. Flat subscription pricing is not sustainable when token consumption per user is scaling unbounded. Usage based pricing shifts the risk back to the buyer, which stabilizes margins but creates the exact budget shock dynamic that drove Uber and Microsoft's own teams to cut licenses.
---
What the 90% Efficiency Gain Number Actually Implies for the Trajectory
One data point from the Tunguz LinkedIn thread deserves more attention than it's gotten: a commenter cited 90% more tokens per GPU delivered within a year. That is not a minor optimization it is the kind of efficiency trajectory that historically reshapes pricing floors across an entire market.
If that rate of improvement holds and it's an open question whether it will the cost per token curve implies that models and inference that today require budget shock level spend could look affordable within 24-36 months. Goldman Sachs' 24x consumption growth forecast by 2030 only makes economic sense if costs compress significantly alongside volume.
The Phi-4 model release Microsoft dropped its small open source models the same day as earnings is a strategic move in this direction. Small, efficient models that run cheaply at inference time are the pressure valve on enterprise budget shock. They're also Microsoft's hedge against the Chinese low-cost inference wave. You can't compete with $0.42/million token output pricing on a GPT-4-class model. You can get closer with a well tuned small model that handles 80% of enterprise workloads at a fraction of the cost.
Whether that strategy succeeds depends on whether enterprise buyers actually trust smaller models for production workflows a question that goes to capability, not just cost.
---
The Strategic Stakes: What Enterprise Buyers and Builders Should Actually Do With This
The operational takeaway from the 100 trillion token milestone is not "AI adoption is working" it's that the unit economics of enterprise AI are at an inflection point where strategy matters more than enthusiasm. Here's what the data actually implies for decisions:
If you're an enterprise buyer: The move to usage based pricing is not a gift it's a risk transfer. Fixed subscriptions gave you cost predictability. Token based pricing gives vendors margin stability and gives you exposure to usage spikes. Audit your AI workflows for token efficiency now, before agentic systems scale that consumption 10x. The companies that survived the 2025-2026 budget shock cycle were the ones who built consumption monitoring into their AI stack before the bills arrived.
If you're a builder or developer: The competitive moat is no longer access to frontier models it's inference efficiency and model selection discipline. A 178x price differential between commodity Chinese models and premium Western frontier models is not a gap you can ignore when you're building at scale. The right answer is probably a tiered model strategy: frontier models for tasks that require frontier capability, smaller or lower cost models for everything else. This isn't a permanent architecture the landscape will shift again but right now, cost per task thinking is the professional standard.
If you're watching the infrastructure layer: Microsoft's $80 billion commitment, alongside Google's ~$85 billion, creates a duopolistic infrastructure moat that is effectively impossible to replicate at equivalent scale. The interesting leverage points are at the software and application layers above that infrastructure where the 90% GPU efficiency improvement makes room for entirely new business models that weren't viable 18 months ago.
The 100 trillion token milestone is a real achievement. The efficiency gains are real. The infrastructure investment is real. But the honest read of the full picture the buried non-AI Azure quote, the Copilot margin hit, the enterprise license cancellations, the Chinese pricing wedge is that the most important work in enterprise AI right now is not about token volume. It's about figuring out which of those trillions of tokens are actually generating economic value proportional to their cost.
That question doesn't have a clean answer yet. Which is exactly why it's the only one worth asking
---
If enterprise AI token costs compress 90% in the next 18 months the same way they did in the last 12 does that vindicate the $80 billion infrastructure bets, or does it mean every company that built a business model on premium token pricing just built on sand?
Loading comments...