Three Moats, One Winner: Why Gemini, Anthropic, and DeepSeek Cannot Both Win

The Triangulation Problem

Three dossiers released this week — Gemini 3.1 Pro, Claude Mythos, DeepSeek V4 — look superficially like three independent capability announcements. Read together, they reveal a structural constraint: each of the three labs has exactly one durable competitive asset, and each asset is structurally preventing acquisition of the other two.

Google's asset is distribution. Gemini reached 750 million users as of March 2026, backed by Workspace, Android, Search, and Vertex AI channels that no competitor can replicate organically. Gemini 3.1 Pro demonstrates top-of-class reasoning (77.1% ARC-AGI-2, 94.3% GPQA Diamond) and unmatched multimodal performance (78.2% Video-MME, leading next-best by 6.8 points). But three months after the preview release on February 19 2026, the model remains in preview rather than general availability. The practical explanation is specific: preview status after 3 months suggests production reliability and latency under load are not yet solved at the 750M-user scale that GA requires. Google's distribution capacity has outrun its serving capacity — the asset is blocking deployment of the capability it owns.

Anthropic's asset is capability. Claude Mythos's UK AISI evaluation (73% expert CTF success, 22-of-32 steps on TLO takeover, 181 successful JavaScript shell exploits vs Opus 4.6's 2) describes a model that is qualitatively ahead of the next-best frontier system. In any other commercial era, that capability gap would translate to pricing power and distribution advantage. Instead, Anthropic has voluntarily confined Mythos to a 40-member Project Glasswing consortium under ASL-4 containment — zero API revenue during peak margin quarters, zero consumer distribution, zero open weights. Anthropic's capability lead is so pronounced that it triggered its own RSP v3.0 containment protocol, making the capability commercially unmonetizable through normal channels. The asset is blocking both distribution and the weights simultaneously.

DeepSeek's asset is open weights. Apache 2.0 licensing (based on R1 precedent and NxCode reporting for V4) means any enterprise, government, or individual can deploy the model locally without API dependency, without per-token costs, without data egress to a US provider. At ~$0.30/MTok projected pricing versus $15/MTok for Claude Opus 4.6 or GPT-5.4, the cost differential is not 10x — it is 50x. But DeepSeek has missed multiple V4 launch windows through March-April 2026, still has not independently validated its 90% HumanEval and 80%+ SWE-Bench claims, and — critically — Reuters reported on April 3 that V4 will run on Huawei chips rather than NVIDIA silicon. For US federal procurement, EU regulated sectors, and most Fortune 500 enterprise supply-chain reviews, Huawei dependency is a non-starter. DeepSeek's open-weight advantage is blocked from the enterprise distribution channel that would make it revenue-positive at scale.

Why Each Moat Blocks the Others

The critical insight is that each lab's moat is ACTIVELY preventing acquisition of the other two moats. These are not independent engineering problems that can be solved in parallel — they are mutually exclusive strategic positions.

Google's distribution problem is the production reliability problem. 750M users means GA requires 99.9%+ SLA that preview testing cannot support. The preview phase is precisely the period where latency, cost, and error-rate scaling under production load get solved. Pushing Gemini 3.1 Pro to GA before these are solved would damage Google's credibility at scale. But the longer Gemini 3.1 Pro stays in preview, the more time competitors have to release capability-matched models with different distribution patterns (Anthropic releasing Mythos API for regulated sectors; DeepSeek V4 GA for open-weight markets). Google's distribution scale creates the production-reliability bottleneck that delays its capability deployment.

Anthropic's capability lead is the safety-containment trigger. RSP v3.0 was written before Mythos existed but was calibrated for exactly this capability tier — autonomous exploitation of novel vulnerabilities at scale. Anthropic did not voluntarily choose Glasswing containment because the company wanted to leave revenue on the table. The containment is a direct consequence of the capability level. If Mythos were released on the public Claude API with standard enterprise safeguards, the company would face immediate regulatory scrutiny (UK AISI evaluation makes ASL-4 claims visible to regulators) and potential liability chains that shadow the White House preemption strategy. Anthropic's capability lead is the exact thing that triggers the safety containment that prevents distribution.

DeepSeek's open-weight commitment is the Huawei dependency. Domestic Chinese chip supply is the only compute path available for training a trillion-parameter MoE outside US export controls. Apache 2.0 licensing is politically inseparable from the domestic-chip narrative — releasing open weights signals China's commitment to post-sanctions autonomous AI development. If DeepSeek pivoted to proprietary licensing and Huawei-independent deployment (using pre-sanction NVIDIA inventory or external compute partnerships), the model would lose its identity as the 'decentralized, sanctions-proof' alternative. The open-weight moat is structurally locked to Huawei dependency by geopolitical position, not by technical choice.

The Winner: Whoever Acquires a Second Moat First

The next 12-month prediction is direct: whichever lab FIRST acquires a second moat wins the 2026-2027 cycle. The three scenarios have distinct probability profiles:

(1) Google solves Gemini 3.1 Pro GA serving (capability + distribution). Probability: medium-high. Google's TPU v6 capacity is the bottleneck, and if TPU v7 shipping in late 2026 resolves it, Google holds reasoning leadership plus 750M users plus Vertex AI enterprise channel. This is the most likely outcome given Google's historical pattern of 6-9 month preview-to-GA cycles. If this happens, Google's dominance in consumer AI becomes locked for 24+ months — Search integration alone is a moat the other labs cannot penetrate.

(2) Anthropic finds a commercial form for Mythos beyond Glasswing (capability + distribution). Probability: low in the short term, high long term. The Glasswing model IS a B2B distribution channel, just a narrow one — it captures ~40 large enterprises at a presumably extreme price point. A 'Mythos Tier' within the normal Claude API, with gated access for verified security teams, would be the economically rational move if the ASL-4 containment can be operationalized. This would convert Mythos from a safety showcase into a revenue-generating moat. Expect 6-12 months for this to materialize, but when it does, Anthropic captures the highest-margin customer segment in enterprise AI (regulated industries, security-critical workloads, insurance-constrained buyers).

(3) DeepSeek V4 launches with verified benchmarks on a non-Huawei path (weights + independent distribution). Probability: low. This would require either the launched V4 running on pre-sanction NVIDIA inventory or a Chinese-chip-compatible deployment tooling stack emerging in US/EU enterprise. Neither is visible in current signal. DeepSeek's only path to a second moat is if geopolitical decoupling slows faster than expected and Huawei chips become acceptable in Western enterprise supply chains. This is plausible for 2027-2028 but unlikely for 2026.

What Each Lab Is Watching in the Others

The strategic tells are precise. Google watches for Anthropic's Mythos-API announcement as the signal that capability is moving to the pricing market. An official Mythos tier — even at stratospheric pricing and narrow access — means Anthropic has solved the monetization problem and is moving toward Google's distribution scale. Anthropic watches for Gemini 3.1 Pro GA as the signal that distribution-scaled reasoning is on the market at $2/MTok. That price point collapses the AI-skills wage premium and enterprise willingness to pay for marginal capability improvements. DeepSeek watches for whether Western enterprise procurement policies relax on Chinese chip dependency — they will not, absent a major geopolitical reset or sanctions relief.

What This Means for Practitioners

For enterprise procurement teams: the next 12 months will clarify which lab holds two moats. If Google reaches Gemini 3.1 Pro GA in Q3 2026, lock in long-term Vertex AI commitments — the pricing and capability combination will be durable for 2+ years. If Anthropic announces Mythos API access before then, prepare for a segmented market where commodity reasoning (Gemini, GPT-5.4, DeepSeek) separates from premium capability (Mythos for regulated/security sectors). If DeepSeek V4 GA with verified benchmarks ships without Huawei dependency, the open-weight layer of AI becomes mission-critical for cost control — but this scenario has low near-term probability.

For ML engineers choosing between platforms: avoid long-term architecture decisions until you see which lab acquires the second moat. If you are building consumer-facing products, assume Google Gemini 3.1 Pro reaches GA in Q3 2026 — design for that integration window. If you are building enterprise security tools, assume Anthropic releases Mythos API on a gated-access basis by Q4 2026 — engineer for both Claude Opus and Mythos capability tiers. If you are building for cost-constrained markets, DeepSeek V4 open weights remain the fallback option despite Huawei dependencies, but do not overcommit to a supply chain dependent on geopolitical resolution.

The contrarian view is worth noting: the triangulation problem may be stable. Oligopolies with orthogonal moats can coexist profitably for extended periods — think AWS/Azure/GCP in cloud. If Google locks in consumer and Workspace distribution, Anthropic locks in regulated-industry safety tier, and DeepSeek locks in the open-weight global south / non-aligned markets, the three-moat stalemate may persist for 2-3 years. But the compounding advantage of holding two moats is so large (distribution-cost synergies, capability-pricing power, weights-ecosystem flywheel) that the first entrant to acquire a second moat likely reorders the entire competitive map within a single quarter.