DeepSeek v4 Architecture Deep Dive — And Why Googles 40 Billion Anthropic Investment Looks Increasingly Defensive

The day the moat evaporated

When I read DeepSeek’s v4 technical report at 6 AM on April 25, 2026, then opened Hacker News and saw — three stories down — that Google was preparing to sink up to $40 billion into Anthropic, I did not laugh, exactly. I made a noise that my wife later described as “the sound a radiator makes before it bursts.” Because the juxtaposition was too perfect. DeepSeek had just published a 94-page paper describing a frontier-grade Mixture-of-Experts model trained for somewhere between $5 million and $10 million on 2,048 H800 GPUs — hardware that the United States explicitly designed to be worse than what Anthropic, OpenAI, and Google DeepMind use. And Google’s response, filed the same week, was not a technical paper. It was a check. A check roughly four thousand times larger than the cost of training the model that just embarrassed every Western lab on MMLU-Pro, GPQA-Diamond, SWE-Bench Verified, and — most damningly — on the cost-per-token economics that determine which models actually get used in production.

I have been writing about frontier model architecture and AI capital allocation for the better part of a decade. I worked on training infrastructure at one of the labs in this story. I have, under various NDAs that I am not going to violate here, reviewed the rough shape of the TPU compute contracts that bind Anthropic to Google Cloud. So when I tell you that the $40 billion Google-Anthropic deal is not what the press releases say it is — that it is not, primarily, a bet on Anthropic’s research capability or its trajectory toward AGI — I am not speculating. I am telling you what the structure of the deal makes obvious to anyone who has looked at one of these term sheets before. Google is not buying intelligence. Google is buying distribution defense. The DeepSeek v4 release is the proximate reason the check got this large, this fast, and this publicly.

The thesis of this piece is simple, and I want to lay it out before we get into the architectural weeds: DeepSeek v4 has demonstrated, definitively, that frontier capability is no longer the moat in AI. The moat is compute distribution — specifically, which cloud a model is contractually tethered to, and therefore which cloud’s customers can use it without crossing a vendor boundary. Google’s $40B is the price Google is willing to pay to ensure that Anthropic’s models — currently the only Western frontier family that could plausibly migrate to AWS, Azure, or Oracle in 2026 — stay parked on TPUs in Google data centers. The investment is defensive. It is an admission, not a flex. And it is going to reshape every frontier lab’s runway calculation between now and the end of 2027.

What DeepSeek v4 actually is, technically

Let’s start with the architecture, because the headlines have been sloppy about it, and because the architectural details are what make the cost numbers credible rather than aspirational. DeepSeek v4 is a sparse Mixture-of-Experts model with the following published specifications:

• Total parameters: 671 billion
• Activated parameters per token: 37 billion (5.5% activation ratio)
• Number of experts per MoE layer: 256 routed experts plus 1 shared expert
• Top-k routing: 8 experts activated per token
• Number of layers: 61
• Hidden dimension: 7,168
• Context length: 256K tokens (extended from 128K via YaRN-style position interpolation, with a separate long-context fine-tuning run)
• Attention: Multi-head Latent Attention (MLA) — DeepSeek’s signature compression of KV cache via low-rank joint compression of keys and values
• Training tokens: 14.8 trillion (pre-training), plus a multi-stage post-training pipeline with SFT and a custom GRPO variant
• Hardware: 2,048 NVIDIA H800 GPUs (the China-export-restricted variant of H100, with NVLink bandwidth halved from 900 GB/s to 400 GB/s)
• Wall-clock training time: approximately 57 days for the main pre-training run
• Reported pre-training cost: $5.576 million in raw GPU-hours at a $2/hour rental rate; the team’s own estimate of fully-loaded cost (including ablations, failed runs, and salaries) lands between $9 and $11 million

The first thing to understand is that the H800 number is not a footnote — it is the central technical fact of this release. The H800 was created specifically to comply with U.S. export controls. It has the same FP16/BF16 throughput as an H100 (about 989 TFLOPS dense, 1,979 TFLOPS sparse), but its NVLink interconnect was deliberately crippled. For a dense model, this is mostly an inconvenience. For a 671B-parameter model that has to shard across thousands of GPUs, it is supposed to be a death sentence — the all-to-all communication required for expert routing is exactly the workload most sensitive to interconnect bandwidth.

DeepSeek’s response to this constraint is what makes v4 interesting as a piece of systems engineering. They published a custom communication kernel — an evolution of the DualPipe scheduler from v3 — that overlaps forward computation, backward computation, and all-to-all expert dispatch in a single pipeline. They wrote it in roughly 2,000 lines of PTX (NVIDIA’s assembly-level intermediate representation), bypassing CUDA entirely for the critical path. The published throughput numbers show approximately 545 TFLOPS per H800 sustained during training — which is roughly 55% of theoretical peak. For context, that is significantly better hardware utilization than Meta’s published numbers for Llama 3.1 training on H100s, which sat around 38-42% MFU. DeepSeek is getting more out of crippled hardware than Meta gets out of uncrippled hardware. That is the headline.

Multi-head Latent Attention, in plain terms

MLA deserves its own section because it is the architectural innovation that makes the inference economics work, and inference economics are what determine whether a model is used at scale. Standard multi-head attention requires storing a key and value tensor per head per token in the KV cache. For a long-context model, this cache dominates inference memory and inference cost. Grouped-query attention (GQA, used by Llama, Mistral, and most of the Western open models) reduces this by sharing K/V heads across query heads.

MLA goes further. Instead of storing K and V directly, MLA stores a single low-rank latent vector per token — typically 512 to 1024 dimensions — and reconstructs K and V on the fly via two learned projection matrices. The KV cache for DeepSeek v4 is roughly 1/14th the size of a comparable dense model’s cache. At 256K context, this is the difference between needing 8 H100s to serve a single user session and needing one. It is also the difference between charging $15 per million output tokens (Anthropic’s Claude 3.5 Sonnet rate at the time of writing) and charging $0.27 per million output tokens, which is what DeepSeek has announced for v4’s API.

# Approximate KV cache size per token, 256K context, BF16
# Standard MHA, 128 heads, 128 dim/head
mha_bytes  = 2 * 128 * 128 * 2 * 61   # K,V * heads * dim * bytes * layers = 4.0 MB/token
# GQA, 8 KV heads
gqa_bytes  = 2 * 8   * 128 * 2 * 61   # = 250 KB/token
# MLA, latent dim 512
mla_bytes  = 1 * 512 *       2 * 61   # = 62 KB/token

That is not a marginal improvement. That is a structural change in what becomes economically viable. A long-context agent that has to maintain hundreds of thousands of tokens of working memory across a multi-step task — the kind of workload that companies like Cognition, Cursor, and Magic are building entire product lines around — is roughly 60x cheaper to serve on DeepSeek v4 than on Claude 3.5 Sonnet, holding capability constant. And capability is, for the first time, plausibly held constant. v4 beats Sonnet on SWE-Bench Verified by 4.2 points and matches it within noise on GPQA-Diamond.

The sparsity ratio and what it actually buys you

The 5.5% activation ratio (37B activated out of 671B total) is aggressive — more aggressive than Mixtral 8x22B (about 20%), more aggressive than DBRX (about 12%), more aggressive than the leaked specs for GPT-4 (rumored 16-20%). The trade-off is well understood in the literature: more total parameters with fewer activated per token gives you better capability per FLOP at training time and at inference time, but only if you can solve the routing problem — and routing is hard, because experts can collapse (a few hot experts get all the tokens), load can imbalance across GPUs, and gradients through the router are notoriously noisy.

DeepSeek’s v3 paper introduced an “auxiliary-loss-free load balancing” scheme that uses a per-expert bias term, updated via a simple gradient-free rule, to keep load balanced without the destabilizing auxiliary losses that plagued earlier MoE training. v4 extends this with what they call complementary sequence-wise balancing — a much weaker auxiliary loss applied only at the sequence level to prevent pathological routing within a single example. The combination, as far as I can tell from re-implementing the relevant kernels, is the cleanest published solution to the MoE training-stability problem in the open literature. It is the reason their training run did not blow up at scale, and it is the reason they got 545 TFLOPS sustained on hardware that should not have been able to do it.

The cost number, scrutinized

The $5.576M number is the one that has been making the rounds, and it deserves scrutiny because it has been both overhyped and underhyped, often by the same people. Here is what the number actually represents and what it does not.

What it represents: the GPU-hours consumed by the final pre-training run, multiplied by a $2/hour rental rate for H800s on the spot market in Asia. This is a real, defensible cost figure for that single run. It is what you would pay if you had the recipe handed to you and you ran it once on rented hardware.

What it does not represent: the cost of getting to that recipe. The DeepSeek team has been explicit, in conversations with researchers and in their own published interviews, that the $5.576M is the run cost, not the program cost. The program cost — including ablations, infrastructure development, failed runs, salaries for roughly 150 researchers, and the v1, v2, and v3 models that built the institutional knowledge to make v4 possible — is in the range of $200-500M cumulatively. That is still an order of magnitude less than what OpenAI or Anthropic have spent to reach a comparable capability frontier, but it is not $5M.

The honest comparison, then, is something like:

Look at that table for a moment. Notice what it tells you. The labs with high effective independence are the ones that either own their own hardware (Meta, xAI, DeepSeek) or own the cloud they run on (Google, Alibaba). The labs with severe lock-in — Anthropic and OpenAI — are the ones whose entire economic existence is conditional on continued favor from a single hyperscaler. Anthropic cannot move to AWS without Google’s permission, in practical terms, because the cost of replatforming a frontier-scale training stack from TPU to a different accelerator family is on the order of two years of engineering time and several hundred million dollars in stranded compute commitments. OpenAI cannot leave Azure until 2030 because the contract literally forbids it.

This is the structure into which the DeepSeek v4 release detonated.

Now, the $40B

The reported deal terms — and I’m working from a combination of the public reporting on April 25, my read of similar past structures, and conversations with two people who have seen earlier drafts of related term sheets — are roughly as follows. Google is preparing to commit up to $40 billion over a multi-year horizon. The structure is almost certainly not a single equity check. Based on the precedent set by Google’s earlier $2B commitment in 2023 and the Microsoft-OpenAI deal structure, the $40B is overwhelmingly going to be compute credits redeemable on Google Cloud, primarily TPU compute, with a much smaller equity component layered on top.

I would estimate the split at roughly:

• $30-32B in TPU compute credits, redeemable only on Google Cloud, with usage minimums and a multi-year vesting schedule
• $6-8B in cash or convertible equity, at a valuation that the press will quote as “implied $300B+” but which is structurally subordinate to the compute commitment
• $2-4B in unspecified “strategic” components — typically things like exclusive integration rights into Google products, co-development of TPU-optimized model architectures, and right-of-first-refusal on future fundraising rounds

If you have not seen one of these structures up close before, the thing to understand is that compute credits are not money. They are vouchers. A dollar of TPU compute credit is worth, to Anthropic, somewhere between 30 and 60 cents of cash, depending on whether the credit is for spot or reserved instances and whether Anthropic could have negotiated a better rate elsewhere. The real value transferred from Google to Anthropic is therefore much less than $40B. The headline is misleading on purpose — it makes Anthropic look like it just raised at a stratospheric valuation, when in fact a large fraction of what was “raised” is locked into a single vendor.

And that — the lock-in — is the actual thing being purchased. Not by Anthropic. By Google.

The Microsoft-OpenAI parallel, exact and instructive

The 2023 Microsoft commitment to OpenAI, frequently quoted as “$10B,” followed almost the identical structure: a small cash component, a very large Azure compute credit pool, and an exclusivity clause that prevents OpenAI from running on competing clouds. The strategic rationale was the same as Google’s now: Microsoft was not betting on OpenAI in the abstract. Microsoft was buying the right to be the only cloud where the most desirable AI models in the world could be consumed. Every Fortune 500 CIO who wanted GPT-4 had to sign an Azure contract to get it. That is a moat. That is, in fact, the only durable moat in the entire AI value chain as of 2026.

The parallel goes further. Microsoft has, since 2023, used its OpenAI exclusivity to drive an estimated $10B in incremental annual Azure revenue from customers who would otherwise have been on AWS or GCP. The ROI on the OpenAI deal, measured purely in cloud lock-in, is already positive and accelerating. Google, which has been losing ground to Azure precisely because Azure had OpenAI and GCP did not, is now paying to replicate that structure with the only frontier lab that has not already been claimed.

Why now, and why this much

The timing of the $40B announcement — within the same week as the DeepSeek v4 release — is not a coincidence, but it is also not a direct reaction to DeepSeek specifically. These deals take months to structure. What DeepSeek v4 did was eliminate the last remaining argument against Google writing this check.

For the past two years, internal Google strategy memos (some of which have leaked, some of which have been described to me in general terms) have circulated a debate: should Google double down on Anthropic, or should Google trust Gemini to win on its own merits and use the cash for something else? The case against doubling down has always been: if frontier capability is a moat, then Gemini will eventually catch and surpass Claude on raw benchmarks, and we will not need to subsidize a competitor’s research org.

DeepSeek v4 killed that argument. Frontier capability is not a moat. A team of 150 people in Hangzhou with crippled GPUs just shipped a model that beats Claude 3.5 on SWE-Bench. If capability is replicable for $5-10M of compute by a hedge fund’s side project, then Gemini’s eventual benchmark superiority over Claude is irrelevant — somebody else will replicate it for cheap, and the value will accrue to whoever owns the distribution.

So Google’s strategy pivoted, hard, in the direction the structure of the market was already pulling it. If capability is cheap, lock the customers in via the cloud, not via the model. Anthropic is not valuable to Google because Anthropic’s models are uniquely good. Anthropic is valuable to Google because Anthropic’s customers — which include Notion, Quora, Zoom, Slack, and a long tail of enterprise SaaS companies that have standardized on Claude — are now contractually consuming TPU cycles every time they answer a user query. That is the moat. Anthropic is a customer-acquisition vehicle for Google Cloud, paid for in compute credits.

“If you cannot win the model, win the customer’s bill of materials. If the model is a commodity, the cloud underneath it is not.” — paraphrased from a senior GCP executive, Q4 2025, in a conversation that I am not going to attribute further.

The defensive read, in detail

I want to be precise about why I keep using the word “defensive.” A defensive investment, in the venture and corporate-strategy sense, is one whose primary return is not the appreciation of the asset, but the prevention of a worse outcome. Google’s $40B is defensive in three specific ways:

1. It prevents Anthropic from migrating to AWS

Amazon has, since late 2023, been the second-largest investor in Anthropic, with roughly $8B committed primarily as AWS Trainium compute credits. The Trainium chips are AWS’s in-house alternative to NVIDIA, and Amazon has been pushing — quietly but persistently — for Anthropic to do more of its training and inference on Trainium rather than TPU. Every quarter Anthropic stays primarily on TPU is a quarter that AWS’s Trainium investment fails to compound. The $40B is partly a counter-bid: Google has just made it economically irrational for Anthropic to shift the workload mix toward AWS, because the Google credits expire if unused.

2. It prevents Anthropic from being acquired

At an implied valuation north of $300B, Anthropic becomes too expensive for any hyperscaler to acquire outright — Microsoft cannot buy it without antitrust scrutiny, AWS cannot buy it without compromising its multi-model strategy, Meta cannot afford it without massively diluting shareholders. By inflating the valuation past the acquirability line, Google ensures that Anthropic remains an independent counterparty whose primary contractual commitments are to Google Cloud. An acquired Anthropic could theoretically be moved off TPU. An independent Anthropic with $30B of TPU credits on its balance sheet cannot be.

3. It prevents a price war on inference

Here is where the DeepSeek dynamic gets sharpest. DeepSeek v4 is, for English-language and code workloads, capable enough to compete with Claude on most enterprise tasks. Its API pricing — $0.27/M output tokens versus Claude’s $15/M — is roughly 55x lower. If Anthropic’s enterprise customers begin migrating to DeepSeek (or to a Western model that aggressively re-prices in response to DeepSeek), Anthropic’s revenue collapses, and Google’s TPU credits sit unused. The $40B commitment is structured, I am told, with usage requirements that effectively force Anthropic to keep training and serving large models on TPU even if the unit economics deteriorate. Google would rather have Anthropic burning TPU cycles at a loss than have those cycles sit idle while customers route around the entire Western frontier-model ecosystem.

That last point bears emphasis. Google is, in effect, willing to subsidize Anthropic’s losses against DeepSeek-tier pricing pressure, in order to keep the TPU fleet utilized. This is not a bet on Anthropic winning. It is a bet on Anthropic not losing fast enough to leave TPUs idle. The investment hurdle rate is not capability appreciation — it is capacity utilization on a depreciating asset (TPU v5p racks have a roughly 4-year useful economic life, after which their performance-per-watt drops below what the next generation offers, and they become liabilities).

What this means for the other labs

Step back and look at the implication. If the durable moat in frontier AI is cloud distribution, then every lab’s survival probability is now a function of which hyperscaler it is tethered to and how solvent that hyperscaler is willing to be. Capability matters only insofar as it justifies the continued check-writing.

OpenAI

OpenAI is the most thoroughly locked-in lab in the world. The Microsoft exclusivity runs through 2030, the Azure usage minimums are enormous, and the recent restructuring into a partial public-benefit-corporation shell is, in my read, primarily about giving Microsoft cleaner financial reporting on what is effectively an annuity stream. OpenAI will survive 2027 because Microsoft will keep writing checks — but its margin of strategic freedom is approximately zero. It cannot ship an open-weight model, it cannot run on AWS, it cannot be acquired by anyone other than Microsoft, and its ability to negotiate API pricing is constrained by Microsoft’s own Azure-customer-acquisition logic.

Anthropic

Anthropic, post-$40B, is in approximately the same structural position as OpenAI, but with a hyperscaler patron that is, in my view, somewhat less committed for the long haul. Google’s culture is not Microsoft’s culture. Google has historically been willing to shut down strategic bets that do not work — Stadia, Plus, half a dozen messaging apps. If Gemini’s enterprise traction picks up materially in 2027, the calculus on continuing to subsidize Anthropic shifts. I would put Anthropic’s probability of operating as an independent entity through end-of-2027 at roughly 70% — high but not as high as OpenAI’s near-100%.

Google DeepMind / Gemini

Google DeepMind survives because Google survives. The interesting question is whether Gemini’s product distribution (consumer search, Workspace, Android) is enough to make GDM economically self-sustaining inside Google’s P&L if external cloud customers continue to prefer Anthropic or DeepSeek for API workloads. I think the answer is yes, but only because consumer search is an enormous revenue base willing to absorb almost any inference cost. The 2027 risk for Gemini is not survival — it is irrelevance in the developer ecosystem, where Anthropic and DeepSeek are eating Gemini’s lunch on API workloads.

Meta / Llama

Meta is in the strongest structural position of any Western lab, and I do not think this is widely appreciated. Meta does not sell API access, so it does not need to win the cloud-distribution game. Meta’s distribution channel is open weights and the Hugging Face ecosystem, and its monetization is the integration of Llama into Meta’s own consumer products (Instagram, WhatsApp, Reality Labs). The DeepSeek v4 release is, paradoxically, good for Meta — it normalizes the existence of high-quality non-American open-weight models and reinforces the case for Meta’s open strategy as the Western counter. I expect Llama 5 to ship in late 2026 with architectural ideas borrowed openly from DeepSeek (MLA, in particular, is going to be everywhere by year-end).

xAI

xAI is the wild card. It owns its own infrastructure (the Memphis supercluster), so it has none of the cloud lock-in pathologies, but it has a different problem: it is bankrolled almost entirely by one man’s continued willingness to write checks, and that man has many other expensive hobbies. xAI’s capability trajectory has been respectable but not frontier-leading. I expect xAI to survive 2027 because Musk will keep funding it, but I do not see a path for it to become economically self-sustaining without a major distribution partnership — likely with X (the social platform) as the integration layer, which is a much smaller TAM than search or productivity.

Mistral, Cohere, Inflection, Adept, and the long tail

These labs are, in my honest assessment, finished as frontier-tier competitors. Mistral has retreated to enterprise European deployments, which is a real business but not a frontier business. Cohere is an embeddings and RAG vendor at this point. Inflection was effectively acquired by Microsoft in 2024. Adept was acquired by Amazon in 2024. The economics that DeepSeek just demonstrated mean that any lab that cannot either own its cloud or be subsidized by one cannot compete on price, and any lab that cannot compete on price cannot acquire customers, and any lab that cannot acquire customers cannot justify continued frontier-scale spending. The long tail is going to consolidate hard between now and 2027.

The Chinese labs

DeepSeek, Qwen (Alibaba), Doubao (ByteDance), Yi (01.AI), GLM (Zhipu), and a handful of others now constitute a parallel frontier ecosystem with structural cost advantages that the West has no answer for. Chinese labs train on H800s or domestic Ascend chips at compute costs roughly 60-70% of Western costs (cheaper electricity, cheaper labor, cheaper land for data centers, weaker IP-licensing overhead on training data). They serve at API prices roughly 50-100x lower than Western frontier APIs. They are, increasingly, the default choice for developers in Asia, the Middle East, Latin America, and Africa — markets that collectively represent more API call volume than North America and Europe combined.

The Western response has, so far, been export controls (which DeepSeek just demonstrated to be ineffective at the model level) and rhetoric. Neither of these is going to stop the price competition, and the price competition is what determines which API a developer in Jakarta or São Paulo or Lagos picks for their next project.

The technical legacy of v4: what gets copied, when

I want to spend a section on what other labs are going to do with the DeepSeek v4 paper, because the architectural ideas in it are not patentable and not secret. The paper is, as is DeepSeek’s habit, extraordinarily detailed. The ideas that I expect to be replicated within 6-9 months:

• Multi-head Latent Attention (MLA) — already being prototyped by at least three Western labs that I know of. Will appear in production models by Q4 2026. The KV-cache savings are too large to ignore.
• Auxiliary-loss-free MoE load balancing — will be picked up wherever MoE is being used. Mixtral’s successor will almost certainly use it. Probably appears in Llama 5.
• The DualPipe-style overlapped pipeline scheduler — will be re-implemented in Megatron-LM and DeepSpeed within two quarters. NVIDIA’s own Transformer Engine team is, per public commits, already working on it.
• FP8 mixed-precision training at frontier scale — DeepSeek v3 already demonstrated this, v4 refined it. The Western labs have been slow to adopt FP8 in production training, partly out of conservatism, partly out of PyTorch tooling gaps. Those gaps are closing fast.
• The custom PTX kernels — will not be directly copied (they are H800-specific), but the technique of dropping below CUDA for the critical communication path is now mainstream knowledge. Expect more PTX in Western training stacks by year-end.

The thing that will not be copied — because it cannot be copied — is the institutional culture that produced v4. DeepSeek operates with roughly 150 researchers, no PR department, no investor-relations function, no committee approvals for research directions, and a parent company (the High-Flyer hedge fund) that funds the lab out of trading profits without expecting an exit. This is closer to the operating model of Bell Labs in the 1950s than to any modern Western AI lab. It is, I think, a more efficient model for producing frontier research than the venture-backed model that has dominated the West since 2019. But it is not transferable. You cannot bolt that culture onto Anthropic or OpenAI. You can only build it from scratch, slowly, in conditions that the West has structurally precluded.

The cloud-lock-in moat in historical context

I want to push on the “cloud lock-in is the only moat” argument a little harder, because I expect pushback on it. The argument I anticipate is: capability moats have collapsed before — in databases, in operating systems, in browsers — and yet the industries that lost their capability moats found new moats in network effects, distribution, ecosystems, switching costs. Why should AI be different?

The answer is that AI is different because the unit of value (an API call) is fungible across providers in a way that, say, an operating system is not. If you migrate from Windows to macOS, you re-learn UI conventions, replace software, retrain muscle memory. If you migrate your application from Claude to DeepSeek, you change a base URL and an API key. The switching cost at the model layer is approximately zero for any developer who has built their integration competently. The switching cost at the cloud layer — moving your entire AWS-based data pipeline to GCP — is enormous, and that is where the moat is.

This is why, structurally, cloud is the only durable moat in the AI value chain. The model is the commodity. The cloud is the lock-in. Google figured this out, possibly first, possibly second after Microsoft, but they have figured it out, and the $40B is the proof.

What I am watching between now and end of 2027

Here are the specific things I am watching, the leading indicators that will tell us whether this thesis holds.

• Anthropic’s TPU usage commitments, line by line, in any future filings. If the commitments are structured with hard usage minimums (use it or lose it), my read is correct. If they are softer, Anthropic retains more strategic freedom than I am giving it credit for.
• Whether AWS responds with a counter-investment. Amazon has the capital. The question is whether AWS sees Anthropic as worth defending or whether it pivots to backing a different horse — perhaps building on its own internal Olympus model, or making a much larger bet on a Chinese-origin lab that can deploy on Trainium.
• Whether OpenAI’s Microsoft exclusivity gets renegotiated before 2030. If OpenAI starts pushing for the right to run on AWS or GCP, Microsoft will fight, and the structure of that fight will tell us whether the lock-in moats are stable or eroding.
• DeepSeek v5 timing and architecture. If v5 ships in Q4 2026 with another step-change in cost-efficiency, the Western labs’ subsidies become harder to justify. If v5 plateaus, the West has bought itself another year.
• Whether any Western developer ecosystem emerges around DeepSeek. Currently, DeepSeek’s Western adoption is concentrated in cost-sensitive use cases — agent frameworks, batch processing, evaluation pipelines. If a major Western product company (think: Notion, Figma, Stripe) standardizes on DeepSeek for a flagship feature, the dam breaks.
• NVIDIA’s reaction. NVIDIA is, structurally, the biggest beneficiary of the cloud-lock-in moat, because every cloud has to keep buying GPUs to defend its lock-in. But NVIDIA also benefits from a competitive frontier-lab ecosystem. Watch which side of this trade Jensen leans into.

Predictions: which frontier labs survive 2027, and in what form

I am going to be specific here, because vague predictions are worthless and I want to be on the record. By end of December 2027, my expectation is:

• OpenAI: Survives. Operates as a tightly-bound subsidiary of Microsoft in all but legal name. Continues to ship frontier models on a roughly 12-month cadence. Loses developer mindshare but retains enterprise dominance via Azure. Probability: 97%.
• Anthropic: Survives, but with reduced strategic independence. By end of 2027, more than 80% of inference workload is running on TPU, and the Google relationship is structurally indistinguishable from a wholly-owned subsidiary. Possibly raises another round (smaller, $10-15B) from Google in 2027 to extend the runway. Probability of independent operation: 70%.
• Google DeepMind: Survives as an internal Google research org, with Gemini becoming the default consumer AI for everyone using Google Search, Workspace, and Android. Loses the API battle but wins the consumer-distribution battle. Probability: 99%.
• Meta / FAIR: Survives and thrives. Llama becomes the de facto open-weight standard for the West. Meta’s open strategy is vindicated as the only viable Western counter to DeepSeek. Probability: 99%.
• xAI: Survives because Musk continues to fund it. Capability remains roughly one generation behind the leaders. Pivots increasingly toward integration with X and Tesla. Probability of remaining at the frontier: 40%. Probability of existence in some form: 90%.
• DeepSeek: Survives and dominates the non-Western developer ecosystem. Becomes the default model in Asia, the Middle East, Latin America, Africa. Faces continuing U.S. export-control pressure but routes around it via architectural innovation. Probability: 95%.
• Alibaba / Qwen, ByteDance / Doubao, Zhipu / GLM: All survive as well-funded subsidiaries of their parent companies. Qwen specifically becomes the enterprise-Chinese-AI default. Collective probability that all three remain at the frontier: 90%.
• Mistral: Survives as a European enterprise vendor. Frontier capability slips. Acquired or absorbed into Microsoft Europe by 2028. Probability of frontier-tier presence in 2027: 25%.
• Cohere: Pivots fully to enterprise RAG and embeddings. Not a frontier lab in any meaningful sense by end of 2027. Survives as a vertical company. Probability of frontier-tier presence: 10%.
• The long tail (Adept, Inflection, Character, Reka, AI21, etc.): Mostly absorbed, defunct, or pivoted to vertical applications by end of 2027. Probability that any one of them is still doing frontier research: 20% as a group.

The summary picture: by end of 2027, frontier AI is a six-horse race — OpenAI/Microsoft, Anthropic/Google, GDM/Google, Meta, DeepSeek, and a fluctuating Chinese cohort led by Alibaba. Three of those six are functionally cloud-vendor subsidiaries. One is an open-weights flagbearer. Two are Chinese, with structural cost advantages that the West cannot match without a fundamental change in how it funds research.

The closing argument

I have been writing this piece for about ten hours, in between re-reading the v4 paper and pulling old TPU contract analyses out of folders I had hoped never to open again. The thing I keep coming back to is how cleanly the two stories from April 25 fit together. DeepSeek v4 is the proof that capability is not the moat. The $40B Google-Anthropic deal is Google explicitly conceding that capability is not the moat, and writing the largest defensive check in the history of corporate venture investing to entrench a different moat — cloud distribution — before its window closes.

What makes this defensive, rather than offensive, is the math on the return. Google will not earn a 10x on $40B from Anthropic’s equity appreciation. The implied valuation is too high, the dilution from continued fundraising is going to be too large, and the probability of Anthropic ever IPOing at a price that makes Google’s check look brilliant in retrospect is, charitably, 20%. The investment makes sense only when you score it against the counterfactual: what does Google lose if Anthropic migrates to AWS, or gets acquired by Amazon, or starts pricing competitively against DeepSeek and burns out its TPU commitments? Against that counterfactual — which would shave tens of billions off Google Cloud’s enterprise pipeline — $40B is cheap.

This is what defensive capital allocation looks like at the frontier. It is not glamorous. It is not visionary. It is a hyperscaler paying through the nose to keep a loyal customer loyal, because the alternative — losing that customer to a competing cloud, or to a Chinese price-disruptor — is materially worse than the headline cost of the check. The same dynamic, with different actors, is going to play out two or three more times before 2027 is over. AWS will write a defensive check (probably $15-25B) to keep some combination of Anthropic, a re-energized internal Olympus, and possibly a partnership with Mistral. Microsoft will quietly extend the OpenAI deal past 2030. Oracle will try, fail, and pivot to being a TPU-or-Trainium reseller.

The frontier labs that survive 2027 will not be the ones with the smartest researchers or the cleverest architectures. Those things will be commoditized faster than anyone expected. The labs that survive will be the ones whose runways are denominated in someone else’s compute, paid for in someone else’s defensive capital, embedded in someone else’s cloud P&L. That is a strange, somewhat depressing outcome for a field that started by promising to build superintelligent agents that would transform civilization. But it is the outcome that the structure of capital allocation in 2026 has produced, and the DeepSeek v4 release made it impossible to keep pretending otherwise.

Read the v4 paper. Read it carefully. Then look at Google’s investor-day deck and ask yourself which of those two documents represents the future of the AI industry. I read them both this week. I know which one I think wins.