A new analyst note is rattling the AI investment story just as the biggest cloud companies lock in record spending plans. The Register reported on October 13, 2026 that cheap Chinese and open-source large language models could trigger a trillion-dollar AI infrastructure reset, citing a Jefferies research note that puts roughly $2.3 trillion in contracted AI capacity backlog at Microsoft, Alphabet, Amazon, and Oracle. About half of that backlog, the note reportedly estimates, traces back to commitments tied to OpenAI and Anthropic.
The timing matters. Those same four companies are projected to spend close to $993 billion in combined capital expenditure in 2027, according to the figures cited in the report. That’s a number large enough to reshape global supply chains for GPUs, power, and data center real estate on its own. Now add a second variable: a wave of low-cost Chinese and open-source models, with Kimi K3 singled out as a prime example, that are pushing down what customers pay for AI inference. If token prices keep collapsing while frontier-model quality keeps converging, the economics behind that trillion-dollar buildout get a lot harder to defend.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What The Register’s Report Actually Says
According to The Register’s coverage of the Jefferies note, the core argument is straightforward: hyperscalers have signed contracts and made capacity commitments on the assumption that demand for premium, high-margin AI inference will keep climbing. Chinese open-source models are undercutting that assumption by shipping comparable capability at a fraction of the price. The note includes two pointed lines worth sitting with. One warns that it can no longer be assumed, especially with the arrival of models like Kimi K3, that large language models will ever be profitable given the ongoing collapse in token pricing. The other asks directly whether the four hyperscalers are really going to spend $993 billion on capex in 2027, and suggests the AI story is due for a big reset at some point in the not-too-distant future.
It’s important to be precise about what is confirmed and what isn’t. The $2.3 trillion backlog figure, the 50% OpenAI/Anthropic attribution, and the $993 billion 2027 capex projection are all attributed to the Jefferies note as reported by The Register. The specific claim that this will definitely trigger a reset, rather than merely create the conditions for one, remains a forecast, not a confirmed outcome. The identities of the analysts behind the quoted lines were not disclosed in the available excerpt of the note. Readers should treat the “reset” framing as a risk scenario that a major Wall Street research shop is flagging, not as an event that has already happened.
Why Kimi K3 and Cheap Chinese Models Spook Wall Street
Kimi K3, built by Moonshot AI, is the model Jefferies names directly, and its presence in the note is less about Kimi K3 being the cheapest option on the market and more about what its existence signals. Kimi K3 is actually priced higher than several of its Chinese open-source peers, at roughly $3 per million input tokens and $15 per million output tokens. What spooks analysts isn’t that one number, it’s the breadth of the field around it. DeepSeek V4.1 Flash, Qwen3.8-Flash, and GLM-5.3-Flash are all shipping at a small fraction of that price, and all are credible enough on benchmarks that enterprise buyers are starting to route real production workloads to them instead of defaulting to a premium Western API.
That is the mechanism Jefferies is pointing at. It isn’t that any single cheap model destroys the market overnight. It’s that once “good enough” inference becomes available at a tenth or a hundredth of the price of a frontier model, buyers start asking why they’re paying premium rates for tasks that don’t need frontier-level reasoning. Multiply that re-evaluation across millions of enterprise accounts, and the revenue assumptions baked into trillion-dollar capacity contracts start to look shakier than they did a year ago.
DeepSeek V4.1 Flash vs Kimi K3: The Pricing Gap
The spread between the cheapest and most expensive Chinese models in this category is itself notable. DeepSeek V4.1 Flash is listed at roughly $0.15 per million input tokens and $0.60 per million output tokens during off-peak hours, rising to about $0.30 and $1.20 at peak times. That puts it at roughly a twentieth of Kimi K3’s output pricing. DeepSeek also discontinued its older V4 Pro tier in September 2026, migrating all traffic to V4.1 Flash and billing it at the cheaper Flash rate, a small but telling example of a Chinese lab consolidating around its lowest-cost offering rather than its most capable one. Readers wanting a side-by-side breakdown of DeepSeek’s own tier changes can find more detail in tech-insider.org’s DeepSeek V4.1 Flash API setup walkthrough.
Token Price Collapse, by the Numbers
The pricing data behind this story is the clearest evidence that something structural is happening to AI unit economics. Industry pricing trackers now put the frontier token price index at roughly 84% below its March 2023 baseline, with the median flagship model costing around $6 per million blended tokens across the dozens of models those trackers follow. Chinese open-source and open-weight models sit well below that median, and several have landed at prices that would have been unthinkable for a frontier-grade model just two years ago.
| Model | Developer | Input Price (per 1M tokens) | Output Price (per 1M tokens) | Release Window |
|---|---|---|---|---|
| DeepSeek V4.1 Flash | DeepSeek | $0.15 (off-peak) / $0.30 (peak) | $0.60 (off-peak) / $1.20 (peak) | September 2026 |
| Kimi K3 | Moonshot AI | $3.00 | $15.00 | August 2026 |
| Qwen3.8-Flash | Alibaba | $0.16 | $0.47 | 2026 |
| Qwen3.7-Flash | Alibaba | $0.03 | $0.13 | 2026 |
| GLM-5.3-Flash | Zhipu | $0.15 | $0.50 | 2026 |
| Qwen3-Coder-Next | Alibaba | $0.30 | $1.50 | 2026 |
Put another way: running a benchmark-scale coding task on DeepSeek V4.1 Flash costs roughly $0.27, against close to $1.99 for a comparable task on a premium Western model at maximum reasoning effort, based on the cost comparisons circulating in trade coverage of China’s model wave. That kind of gap is exactly what a $2.3 trillion capex bet is supposed to be insulated against, and it’s exactly what Jefferies is flagging as a risk it currently isn’t insulated against. Readers tracking the broader Chinese open-weight field can compare specs in tech-insider.org’s Kimi K3 vs Qwen3.8-Max vs GLM-5.2 breakdown.
DeepSeek’s January 2025 Shock: A Preview of This Moment
This isn’t the first time a cheap Chinese model has rattled the AI infrastructure story. In January 2025, DeepSeek’s initial breakthrough triggered a sharp selloff in AI-linked stocks, and hyperscaler leadership had to respond in real time. Meta CEO Mark Zuckerberg addressed the moment directly on an earnings call, arguing that “I continue to think that investing very heavily in capex and infrastructure is going to be a strategic advantage over time,” according to Fortune’s reporting on the call.
Zuckerberg also used the moment to frame DeepSeek’s rise as a geopolitical argument for Meta’s own open-source strategy, saying “there’s going to be an open-source standard globally, and I think that for our own national advantage it’s important that it’s an American standard,” per the same Fortune report. He went further on the infrastructure question itself, telling analysts “at this point, I would bet that the ability to build out that kind of infrastructure is going to be a major advantage for both the quality of the service and being able to serve the scale that we want to,” as TechCrunch reported at the time.
That episode is worth revisiting because the hyperscalers won that round. Capex didn’t shrink after the DeepSeek shock, it grew. Microsoft, Meta, and Google combined pushed AI capital spending well past $725 billion by 2026, a figure covered in tech-insider.org’s earlier report on Microsoft, Meta, and Google’s AI capex hitting $725 billion. The question Jefferies is now raising is whether that pattern holds a second time, with a deeper and more price-competitive field of Chinese models than existed in early 2025, and with hyperscaler spending commitments an order of magnitude larger than they were back then.
The $2.3 Trillion Backlog, Broken Down
The numbers in the Jefferies note are large enough that they’re worth laying out plainly rather than folding into prose. The table below isolates the specific figures The Register attributed to the research note, separate from any of the broader market commentary surrounding them.
| Metric | Reported Figure | Source |
|---|---|---|
| Contracted AI capacity backlog (Microsoft, Alphabet, Amazon, Oracle) | ~$2.3 trillion | Jefferies note, via The Register |
| Share of backlog tied to OpenAI and Anthropic commitments | ~50% | Jefferies note, via The Register |
| Projected combined 2027 capex, same four companies | ~$993 billion | Jefferies note, via The Register |
| Microsoft, Meta, Google combined AI capex (2026, reported separately) | $725 billion | tech-insider.org reporting |
| Frontier token price index vs March 2023 baseline | -84% | Industry token-pricing trackers |
The 50% figure is arguably the most important number in the entire note, because it means the backlog isn’t diversified risk spread evenly across every AI customer in existence. Half of it rests on two companies, OpenAI and Anthropic, whose own business models depend on charging enough for inference to eventually turn a profit on the enormous compute they’re renting. If Chinese open-source competition keeps compressing what the broader market is willing to pay for tokens, it puts pressure on exactly the customers whose spending underwrites the largest single slice of the hyperscaler backlog.
Competitive Landscape: US Frontier Labs vs Chinese Open Models
The competitive picture here isn’t a clean two-horse race. On one side sit the frontier labs, OpenAI, Anthropic, and Google, whose latest releases, including Anthropic’s Haiku 5.5 and Google’s Gemini Argon, both reportedly shipped within the past week, compete primarily on capability ceiling and enterprise trust rather than price. On the other side sits a crowded and fast-moving field of Chinese labs, DeepSeek, Alibaba’s Qwen family, Zhipu’s GLM line, and Moonshot’s Kimi, that are competing primarily on price and open licensing, with capability that keeps closing the gap on specific benchmarks even when it doesn’t match frontier models on the hardest reasoning tasks.
Qwen, GLM, and the Long Tail of Cheap Models
What makes this round different from the original DeepSeek shock is the sheer number of credible, cheap alternatives now available. It’s no longer just DeepSeek against the field. Qwen3.8-Flash, Qwen3.7-Flash, Qwen3-Coder-Next, and GLM-5.3-Flash all occupy similar price bands, each backed by a different Chinese technology company with its own cloud distribution and developer tooling. Tech-insider.org’s own comparison of GLM 5.3 Prime, Qwen3.8 Max, and Step 5 preview tracked a benchmark gap of nearly 39 points between the strongest and weakest of that trio, which illustrates how uneven capability still is even within the Chinese model field itself. That unevenness cuts against any simple story that Chinese models are uniformly “good enough.” Some clearly are for narrow, high-volume tasks; others still trail meaningfully on complex reasoning, which is exactly where Western frontier labs are trying to hold their pricing power.
Separately, Nvidia’s own exposure to this dynamic is worth noting. Chip supply to China has itself become a flashpoint, with Nvidia’s share of China’s AI chip market reportedly dropping toward 34% as Huawei’s share climbed to around 28%, according to tech-insider.org’s coverage of Nvidia’s China AI chip share decline. That matters for the capex reset conversation because it shows the compute-supply side of this story is just as contested as the model-pricing side, with Chinese firms pushing to control more of their own AI stack from silicon up through inference pricing.
The Bear Case: Could the Reset Really Happen?
The bear case Jefferies is gesturing at goes roughly like this. Hyperscalers signed long-dated capacity contracts assuming AI inference demand would keep growing at premium prices. If a meaningful share of enterprise workloads can be served just as well by a $0.15-per-million-token Chinese model instead of a $6-per-million-token Western frontier model, the revenue per unit of compute that hyperscalers can extract from AI customers falls. Falling revenue per unit against fixed, already-contracted capacity costs is precisely the setup that forces a write-down, a slowdown in new build-outs, or a renegotiation of existing deals. The note’s own rhetorical question, whether four companies are really going to spend $993 billion in a single year, reads as skepticism that the current spending trajectory survives contact with falling token prices.
There’s also a concentration risk angle. With roughly half the $2.3 trillion backlog linked to OpenAI and Anthropic specifically, any stumble at either company, whether from pricing pressure, a funding gap, or slower-than-expected enterprise adoption, would ripple disproportionately through the hyperscalers that built capacity expecting those two customers to keep scaling.
The Bull Case: Why Hyperscalers Might Keep Spending Anyway
The counterargument, echoing what Zuckerberg told analysts after the original DeepSeek shock, is that infrastructure itself is the moat, regardless of what any single model costs per token. Even if cheap Chinese models capture the lower end of the market, demand for compute used to train next-generation frontier models, to run agentic workloads that chain together millions of calls, and to serve regulated enterprise customers who won’t route sensitive data through a foreign-hosted model, may keep growing fast enough to absorb the planned capacity regardless. Falling token prices can also expand the total addressable market by making entirely new categories of AI application economically viable for the first time, which is itself a demand driver hyperscalers are betting on.
It’s also worth remembering that $2.3 trillion in backlog isn’t the same as $2.3 trillion already spent and stranded. Contracted backlog can be renegotiated, phased, or repurposed across customers and workloads as conditions change, which gives hyperscalers more flexibility than a simple “reset” framing implies.
Historical Parallels: Lessons From Past Infrastructure Overbuilds
Wall Street analysts reaching for the word “reset” are implicitly invoking prior cycles where infrastructure spending outran near-term demand. The dot-com-era telecom buildout of the late 1990s laid enormous amounts of fiber on the bet that internet traffic would justify it; much of that capacity sat unused for years before traffic eventually caught up. Cloud computing itself went through smaller versions of this pattern, with providers occasionally over-provisioning data center capacity ahead of enterprise migration timelines that moved slower than projected.
What makes the current AI cycle different is the compressed timeline. Telecom overbuild played out across most of a decade. The AI capex build-out Jefferies is describing has concentrated roughly $2.3 trillion in commitments within just a few years, and the competitive pressure from Chinese open-source pricing is arriving while the capacity is still being built rather than years after. That compression is exactly why a single analyst note citing one model, Kimi K3, as a symbol of a much larger trend has been enough to reopen the “AI bubble” conversation that had quieted somewhat through much of 2026.
Market Reaction and What Analysts Are Watching
The practical signal to watch in the coming weeks is whether other sell-side research desks echo Jefferies’ framing or push back against it. A single note from one firm doesn’t move markets on its own, but it does set a reference point that other analysts, and company earnings calls, will be measured against. Investors will also be watching whether Microsoft, Alphabet, Amazon, and Oracle adjust their own capex guidance in upcoming earnings reports, or whether they reaffirm the current trajectory the way Meta did after the original DeepSeek shock in January 2025.
Enterprise buyers, meanwhile, are the variable that ultimately decides who’s right. Every workload that migrates from a premium frontier API to a cheaper Chinese open-source alternative is a small data point in favor of the bear case; every new agentic or enterprise deployment that requires frontier-level reasoning is a data point in favor of the bull case. Developers comparing current options across the pricing spectrum can see the range for themselves in tech-insider.org’s Claude Opus 5.5 vs GPT-6 Astra benchmark comparison, which illustrates how close frontier-model capability has become even as pricing between labs diverges sharply.
How Hyperscalers Could Respond
If the pressure Jefferies describes intensifies, hyperscalers have a handful of realistic levers. They can slow the pace of new capacity contracts without canceling existing ones, effectively extending build-out timelines rather than reversing them. They can push harder into price-competitive tiers of their own, something Microsoft, Google, and Amazon have all already started doing with cheaper, flash-tier models of their own. They can lean further into enterprise and government customers who value data residency and compliance over raw price, a segment Chinese open-source models struggle to serve regardless of how cheap their tokens are. Or they can diversify revenue away from pure inference pricing toward higher-margin layers like agent orchestration, specialized tooling, and vertical-specific AI products, where price-per-token matters less than the overall value of the workflow being automated.
Illustrative monthly cost delta (1B output tokens/month):
DeepSeek V4.1 Flash (peak): 1,000 x $1.20 = $1,200
Kimi K3: 1,000 x $15.00 = $15,000
Gap per 1B output tokens: $13,800/month
(Figures based on publicly listed per-million-token output pricing;
illustrative only, not a vendor cost quote.)5 Predictions for the Next Phase of This Story
- More sell-side notes will follow. Jefferies won’t be the last bank to publish a skeptical capex take; expect competing notes from other research shops within weeks, some reinforcing the reset thesis and some defending current spending plans.
- Hyperscaler earnings calls will address it directly. Expect Microsoft, Alphabet, Amazon, and Oracle executives to be asked pointed questions about Chinese open-source pricing pressure on upcoming earnings calls, much as Meta was asked about DeepSeek in early 2025.
- Price tiers will keep splitting. Rather than one uniform reset, expect the market to bifurcate further into cheap, high-volume commodity inference and expensive, premium reasoning tiers, with hyperscalers competing in both.
- Chinese model releases will keep accelerating. With DeepSeek, Qwen, GLM, and Kimi all shipping updates on a roughly monthly cadence through 2026, expect the price-competitive field to widen rather than consolidate in the near term.
- No immediate capex pullback. Based on the pattern after the 2025 DeepSeek shock, the most likely near-term outcome is reaffirmed, not reduced, hyperscaler spending guidance, with any real slowdown, if it comes, showing up gradually across 2027 rather than as a sudden correction.
Frequently Asked Questions
What did The Register actually report?
The Register reported on a Jefferies research note estimating roughly $2.3 trillion in contracted AI capacity backlog across Microsoft, Alphabet, Amazon, and Oracle, with about half tied to OpenAI and Anthropic, and projected combined 2027 capex of about $993 billion for the same four companies.
Has the AI infrastructure reset actually happened yet?
No. The reset described in the Jefferies note is a risk scenario flagged by analysts, not a confirmed event. No hyperscaler has announced a capex pullback tied specifically to this note as of October 13, 2026.
Why is Kimi K3 named specifically in the note?
Kimi K3, from Moonshot AI, is used as a symbol of the broader trend of credible, cheap Chinese and open-source models pressuring token pricing, even though Kimi K3 itself is priced higher than several of its Chinese peers like DeepSeek V4.1 Flash and GLM-5.3-Flash.
How much cheaper are Chinese open-source models than Western frontier models?
Pricing varies widely, but models like DeepSeek V4.1 Flash and Qwen3.7-Flash list output prices well under $1 per million tokens, compared to several dollars per million tokens for many premium Western frontier models, based on current published pricing.
Did this happen before with DeepSeek?
Yes. DeepSeek’s initial model triggered a similar market scare in January 2025. Hyperscalers, including Meta, publicly defended their capex plans at the time, and overall AI infrastructure spending continued to grow afterward rather than shrink.
Which companies hold the largest share of the $2.3 trillion backlog?
The Jefferies note, as reported by The Register, attributes the backlog to Microsoft, Alphabet, Amazon, and Oracle collectively, with roughly half of it tied specifically to contracts and commitments connected to OpenAI and Anthropic.
Could this affect Nvidia and AI chip demand?
It’s a reasonable concern analysts are watching, since slower hyperscaler capex growth would eventually affect chip orders, though no direct Nvidia guidance change has been tied to this specific note as of this report.
What should enterprise buyers take away from this?
The pricing gap between frontier and open-source Chinese models is real and growing, which gives buyers more leverage to match workload complexity to model cost rather than defaulting to the most expensive option available.
Related
- Microsoft, Meta, Google AI Capex Hits $725B [2026]
- DeepSeek V4.1 Flash API Setup: 13 Steps, 1M Tokens [2026]
- Kimi K3 vs Qwen3.8-Max vs GLM-5.2: $10.60 Gap [2026]
- GLM 5.3 Prime vs Qwen3.8 Max vs Step 5: 39-Pt Gap [2026]
- Nvidia China Share Drops to 34% as Huawei Hits 28% [2026]
- Claude Opus 5.5 Edges GPT-6 Astra by 0.9 Points [2026]
Related Coverage
- Dimon: Anthropic's Mythos AI Spikes Cyber Risk 10x [2026]
- OpenAI Engineer, 35, Arrested Over Hidden Office Bug [2026]
- Halluminate's $30M Bet Exposes AI's 51% Finance Gap [2026]
- Character.AI's Age-5 Bankruptcy Shakes AI Rivals [2026]
- Gemini 4 Argon Lands 4 Enterprise Clients at Launch [2026]
