A scenario forecast — by Thorsten Meyer, May 2026
There are six credible Western frontier AI labs in May 2026. By the end of 2028, there will be two, or three, or twelve. Each of these outcomes is internally coherent, supported by different combinations of forces already visible today, and consequential for trillions of dollars of capital allocation.
The question is not which scenario is correct. The question is which one you are positioned for.
This piece is a scenario forecast, not a dispatch. The dispatches in this series cover what happened. This one covers what is about to happen. The method is older than the AI industry — Royal Dutch Shell, RAND, the Hudson Institute. The discipline is to produce internally consistent futures rather than predictions, to identify the forces that branch the futures, and to specify the leading indicators that signal which branch is materializing. The forecast is wrong if it claims certainty. It is useful if it makes the choices visible.
The six labs are Anthropic, OpenAI, Google DeepMind, xAI, Meta Superintelligence Labs, and Reflection AI. The China sphere — DeepSeek, Alibaba’s Qwen, Moonshot AI’s Kimi family, Zhipu — is a parallel ecosystem operating under different capital, regulatory, and compute constraints. The European sphere — Mistral, Aleph Alpha, Black Forest Labs — sits in a third position, structurally aligned with EU regulation and limited compute. The endgame applies to all three spheres but resolves differently in each.
What follows is the structure of the forces, three coherent scenarios for the Western sphere, a tail-risk overlay applicable to all three, and a list of fifteen signposts to watch over the next eighteen months. The piece is longer than my dispatches because the question is bigger.
The 2028 Model Lab Endgame.
How six becomes two, three, or twelve — and which combination of forces decides.
There are six credible Western frontier AI labs in May 2026. By the end of 2028 there will be two, or three, or twelve. Each outcome is internally coherent, supported by different combinations of forces already visible today, and consequential for trillions of dollars of capital allocation. The question is not which scenario is correct. The question is which one you are positioned for.
Six Western labs. Different positions on the same forces.
The competitive picture is easier to compare side-by-side than the financial press has made it. Capital structure, revenue quality, distribution depth, regulatory exposure — each lab sits on a different combination. The same six forces will resolve to different outcomes for each of them.
CORFOTO Fabric 12x10ft AI Research Robot Backdrop Mini Retro Robot with Chemistry Lab Equipment and Blackboard Photography Background for Science Technology Themed Party Photo Props
❤ Size: 12 x 10 ft (W x H) / 140 x 117 in, Horizontal Style
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Six independent forces. Their combinations produce the scenarios.
Each force operates on its own trajectory; the scenarios that follow are simply the three coherent ways the forces can resolve together. None is destiny. All are visible in the data through May 2026.
Compute economics.
Training cost growing 2.4× per year. GPT-4 amortized $40M (2023) → $1B by early 2027 → $10B+ by 2028. Hardware acquisition cost 1–2 OOM higher. Only labs with sustained access to that capital maintain frontier competition.
Capital availability and quality.
Q1 2026: $180B AI funding, more than all of 2024. ~80% to OpenAI, Anthropic, xAI. Sovereign wealth + PE channels dominate. May 4 OpenAI/Anthropic enterprise JV announcements (Blackstone, TPG, Brookfield) confirm: the relationships that matter are with alternative asset managers.
Capability convergence and the open-weight floor.
Stanford AI Index: Chinese frontier “effectively closed” the gap. 3–6 months behind on benchmarks; 1/20th the price per token. Frontier-tier capability is a depreciating asset on a 6–12 month cycle. The model commoditizes; the moat is enterprise distribution.
Talent flow.
$3.4B seed capital to 12 founders departing the major labs in 12 months. xAI lost all 11 co-founders. DeepSeek opening external financing largely to retain talent. The 2027–2028 frontier will be competed for by some of the 6 + 3–5 well-capitalized spinouts + companies not yet founded.
Regulatory gating.
EU AI Act enforcement August 2, 2026. Pentagon two-channel architecture (multi-vendor + Mythos sole-source). Anthropic SCR in litigation. Each lab’s regulatory exposure is now a primary variable in competitiveness.
The agentic transition.
Q1 2026 was the quarter “agentic” stopped being a feature and became a category. May 4 OpenAI/Anthropic enterprise JVs are explicit: forward-deployed engineers, Palantir-style integration, PE-backed channel distribution. Agents are now the unit of economic value, not models.
SQL Server 2025 Unveiled: The AI-Ready Enterprise Database with Microsoft Fabric Integration
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Three coherent futures. One branch point pattern.
The forecast horizon is end of 2028 — long enough for capital cycles to play out, short enough that today’s data points constrain the analysis. The branches fork at three identifiable inflection points: Anthropic’s IPO outcome (Q4 2026), the open-weight capability gap (mid-2027), and the agentic transition’s revenue distribution (Q4 2027).
High-Performance Computing and Artificial Intelligence in Process Engineering
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Each lab. Each scenario. The outcome it implies.
A scenario forecast is only useful if it specifies what each scenario means for each player. The matrix below is the bet you place when you allocate capital. Read across each row to see what happens to a single lab; read down each column to see what each scenario looks like in aggregate.
| Lab · sphere | Scenario A · Duopoly 35% | Scenario B · Equilibrium 30% | Scenario C · Stratification 25% |
|---|---|---|---|
| Anthropic | Scaled · $1.5–2.5TCement duopoly position.Frontier-tier-1 dominant. PE-channel distribution captures enterprise share. Mythos sole-source channel persists. | Tier-1 · $1.2–1.8TOne of three majors.Frontier-tier-1 alongside OpenAI and Google. EU regulated-market share grows; federal SCR situation resolves favorably or expires. | Tier-1 premium · $800B–1.2TAGI-adjacent premium tier.Smaller addressable market; higher margins; revenue concentrated in 5% of workloads requiring genuine frontier-tier-1. |
| OpenAI | Scaled · $1.5–2.5TOther half of duopoly.Microsoft partnership deepens. Conditional Amazon capital arrives in full. PE-channel JV (Development Co) becomes primary enterprise vehicle. | Tier-1 · $1.5–2.0TOne of three majors.Microsoft expands own internal models (Phi-tier) but maintains OpenAI exclusivity for frontier. IPO 2027 at $1.5T+. | Tier-1 premium · $1.0–1.5TAGI-adjacent premium leader.Compute commitments (5GW) become structural overhead; margin compression on commodity workloads. |
| Google DeepMind | Internal supplierCloud-line revenue, not standalone.Frontier capability supplies Google Cloud and Workspace. Not externally measurable as frontier-model business. | Tier-1 · $400–700B notionalThird frontier-tier-1 lab.Cloud growth sustains; AI line item becomes investor-attributable. TPU full-stack matters. | Tier-1 premiumFrontier capability internal.Less commercial differentiation than A or B; consumer-product distribution preserves position. |
| xAI | Defense verticalPentagon Channel 1 specialist.Generalist frontier-tier abandoned. SpaceX IPO is the public vehicle. Federal classified workload concentration. | Sub-frontier · $400–600BSpecialty + Pentagon.Defense-aligned vertical with Musk-network political durability; not frontier-tier-1 generalist. | Tier-2 frontierCommodity-frontier provider.Loses 11 co-founders catches up via SpaceX network; serves federal + Twitter-ecosystem distribution. |
| Meta · Superintelligence | Open-weight exitStops chasing frontier-tier-1.Llama 5 / Muse 2 become open-weight standard; capex revised down; investor pressure forces clarity. | Open-weight enterpriseEnterprise share via cost-efficiency.Open-weight provider of choice for cost-sensitive workloads; sustained capex but disciplined. | Tier-2 frontier · openFrontier-tier-2 leader.Open-weight competition with Chinese cohort; meaningful enterprise share at commodity-tier pricing. |
| Reflection AI | Acquired · $15–25BStrategic capability bolt-on.Microsoft, Google, or Nvidia acquires by mid-2027. Founders cash out; teams integrate. | Persists · $40–80BSpecialty frontier-tier-2.Productization 2026 H2; enterprise customer references signed; possible IPO 2028. | Tier-2 specialistDefense + specialty workloads.Persists at $20–60B; specialization-by-design wins. |
| 12 Founders cohort | 1–2 surviveMost fail or get acquired.Capital crunch compresses options; specialization isn’t enough without distribution. | 3 reach near-frontierThinking Machines, AMI, Periodic.Well-capitalized cohort survives via specialization; 9 fail to scale. | 5–6 viable specialistsVertical specialization wins.Stratification rewards focused capability; 5–6 reach commercial scale. |
| China sphere | Parallel sphereOperating in own zone.3–4 frontier-tier in China; export-controlled access for non-restricted markets; ~3–6 month gap holds. | 4 frontier-tier in sphereStable equilibrium.Gap closes to 3 months; Apache 2.0 base models adopted globally; Alibaba Qwen most-downloaded family. | Tier-2 globallyDefines commodity-frontier.Gap closes to under 3 months; China sphere defines tier-2 pricing globally. |
| Europe sphere | EU-regulated onlyMistral as regional champion.EU Act-driven procurement preference; bounded outside the EU; €30–50B Mistral. | EU + spillover2–3 viable players.Mistral expands beyond EU on cost-efficiency; Aleph + BFL specialize; €40–80B Mistral. | Tier-2 + specialtyModality + sovereign deployment.European bet vindicated as the regulated-market category captures real share. |
Accelerate Everything with Tensor Cores: A Developer’s Guide to High-Performance AI, Efficient Training, and Scalable Models
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
A 15–25% probability event that reshapes any base scenario.
Tail risk is not orthogonal to the base scenarios; it overlays them. Whichever scenario plays out, a Mythos-class capability proliferation event compresses returns, increases regulatory complexity, and shifts the equity structure of the major labs toward government-influenced governance.
The proliferation event that reshapes the equity structure of the labs.
Path 1. A Glasswing consortium member’s access is compromised; nation-state or organized criminal actor obtains Mythos-class capability; major cyberattack on critical infrastructure (financial, power, healthcare). Political response immediate and severe.
Path 2. Open-weight models reach Mythos-class offensive cybersecurity capability independently. Estimated timeline based on capability progression: 12–18 months from May 2026, putting it in 2027 H1–H2 window.
Either path triggers the same response: Defense Production Act authorities, “Strategic AI Reserve” framework with government preferred-equity in Anthropic and OpenAI, mandatory sovereign-cloud deployment for federal-classified workloads. EU does similar via Article 7 reclassification. China closes domestic market.
Probability: 15–25% in 18 months, 30–40% in 36 months. Tail-risk hedging is appropriate in any portfolio with significant frontier-AI exposure. The probability is not low.
Fifteen leading indicators. The next 18 months will tell.
The signposts operate together. A pattern across multiple indicators is more meaningful than any single one. The first six months of EU AI Act enforcement (August 2026 – February 2027) should produce enough signal to identify which scenario is most consistent with the unfolding data.
- Anthropic IPO pricing (Oct 2026). >$1T → A. $700B–$1T → B. <$700B → C or stress.
- OpenAI IPO timing. Announcement before end-2026 → A or B. Delay to 2028 → C or capital stress.
- Meta Q2 capex revision. Pulled back <$115B → B/C. Held or raised >$135B → B.
- Reflection AI productization. Commercial product 2026 H2 → B/C. None by Q1 ’27 → A (acquisition).
- Microsoft positioning. Internal model expansion → B. Deepening OpenAI exclusivity → A.
- Google DeepMind disclosures. Sustained $20B+ Q-over-Q with explicit AI attribution → B viable.
- xAI capability vs SpaceX IPO. Frontier-tier benchmarks before IPO → B. Sub-frontier confirmed → A or vertical-only.
- DeepSeek V5 release. By Q1 2027 at frontier parity → C. Delayed to mid-2027+ → A or B.
- Open-weight gap to frontier. <6mo by end-2026 → C. 9–12mo holds → B. Widens → A.
- Spinout cohort funding rounds. Frontier-tier valuations ($30B+) by end-2026 → B/C. Stalled → A.
- Pentagon multi-vendor expansion. Channel 1 to civilian agencies 2026 H2 → B/C. Consolidation to 2–3 vendors → A.
- EU AI Act enforcement actions. Major US-hyperscaler penalty within 12 months → real teeth (relevant to all).
- Sovereign wealth positioning. Concentration in OpenAI/Anthropic → A. Diversification → B.
- Mythos-class proliferation events. Any major incident or open-weight Mythos-class disclosure → tail risk activates.
- Talent flow direction. Net positive flow to top three → A. Net positive flow to spinouts/tier-2 → B/C.
The endgame is six becoming two, three, or twelve. The bet you place today is the bet on which of those is real.
A note on method
A scenario forecast is not a prediction. It is a description of internally coherent futures that are jointly exhaustive and useful for decision-making under uncertainty. Three scenarios are the canonical number — fewer collapses analysis to false binaries; more diffuses signal. The tail risk is overlaid because crisis-triggered futures are not orthogonal to the base scenarios; they reshape any of them.
The discipline requires that each scenario be:
- Internally coherent: every element of the scenario follows from the same underlying forces
- Causally connected to today: each scenario is a plausible extension of conditions visible in May 2026, not a leap to a possible-but-unsupported future
- Distinct from the others: the scenarios fork at identifiable branch points, with leading indicators that disambiguate
- Strategically consequential: each scenario implies different capital allocation, different organizational positioning, different policy responses
The forecast horizon is end of 2028 — long enough for capital cycles to play out, short enough that today’s data points constrain the analysis. Forecasts beyond 36 months on AI strategy are exercises in narrative, not analysis. This one stays inside the window where the available data still has predictive force.
The probabilities I assign to each scenario are first-order estimates. They are not the output of a model; they are the output of judgment informed by the data assembled in this piece and the prior dispatches in the series. A reader who weights the forces differently will arrive at different probabilities and different scenario rankings. That is appropriate. The value of the forecast is the structure, not the numbers.
I. The terrain in May 2026
The six Western frontier labs sit on capital and capability positions that are easier to compare side-by-side than the financial press has made them.
Anthropic is closing a $50 billion round at a $900 billion valuation, with revenue scaling from $9 billion run-rate at end of 2025 to $30-40 billion by April 2026 — quadrupling in six months. The IPO is scheduled for October 2026. The customer base is enterprise-heavy, with the model favored by senior engineers and regulated industries. Mythos Preview, disclosed April 7, established a single-source defensive cybersecurity capability that the Pentagon is buying outside formal procurement channels. The strategic position is the strongest of any frontier lab on revenue quality, but the company is excluded from the May 1 Pentagon multi-vendor channel, creating a single-channel revenue concentration risk that the IPO disclosure will surface.
OpenAI closed a $122 billion round in March at an $852 billion valuation. The capital structure is unusual: $50 billion from Amazon ($15B upfront cash, $35B conditional on capability and IPO milestones), $30 billion from Nvidia, $30 billion from SoftBank in three $10B tranches through October. The strategic exposure is to the conditional capital — meaningful tranches do not arrive unless OpenAI delivers specific performance milestones, including possibly an IPO by end of 2026. Revenue is approximately $5 billion against $8.5 billion in annualized losses. The compute commitment is 5 gigawatts (3 GW inference, 2 GW training), which is roughly the electricity consumption of a mid-sized European country.
Google DeepMind does not raise external capital because it is internal to Alphabet. The relevant disclosures are Alphabet’s, and they are good. Q1 2026 cloud revenue exceeded $20 billion (+63% YoY), GenAI products grew 800% YoY, the cloud backlog nearly doubled to $460 billion. The full-stack position — TPUs, Vertex AI, Gemini models, Workspace integration — is the most architecturally complete in the industry. The strategic question is not whether Google can compete on frontier capability but whether Google can convert that capability into the same enterprise dominance that Anthropic and OpenAI have established, given Alphabet’s structurally consumer-led product mix.
xAI raised $20 billion in Series E at the start of 2026, then merged corporate interests with SpaceX. The combined entity is positioned for the SpaceX IPO expected mid-to-late 2026, which will be the primary public-market vehicle for xAI exposure. The company has reportedly lost all eleven of its co-founders since 2023, including chief engineer Igor Babuschkin. Total raised through 2026 is approximately $42.7 billion. xAI was included in the May 1 Pentagon multi-vendor channel and inherits SpaceX’s existing federal-government distribution relationships. Capability disclosures lag the other labs and are subject to Musk’s specific framing preferences.
Meta Superintelligence Labs debuted Muse Spark in April after the $14.3 billion Scale AI investment that brought Alexandr Wang in nine months earlier. Capex guidance for 2026 is $115-145 billion, raised from earlier $115-135 billion. Mark Zuckerberg’s Q1 earnings call answer to “where is the ROI” — “that’s a very technical question” — produced a 6% after-hours stock drop on a quarter with revenue up 33% and profits up 61%. Meta has the largest capex commitment of any lab and the weakest disclosure quality on AI ROI specifically. The strategic position is the most uncertain.
Reflection AI raised $2 billion in 2025 at a $6.8 billion valuation, with backing from Nvidia, AMD Ventures, Salesforce Ventures, Inovia Capital, and Radical Ventures. The founding team is ex-Google DeepMind. Training is in progress on a model reportedly at “tens of trillions of tokens” — frontier-class scale. The company was included in the May 1 Pentagon multi-vendor channel, which is the most consequential commercial development for any lab below the OpenAI/Anthropic tier in the past year. Reflection has the most capital efficiency of the six but the least revenue and the most execution risk.
The China parallel sphere runs on different economics. DeepSeek V4 launched April 24 with 1.6 trillion parameters in a Mixture-of-Experts architecture, 3-6 months behind frontier on standard reasoning per the company’s own technical paper, but at approximately 1/20th the price per token of GPT-5.5. DeepSeek is opening a $20 billion funding round, which would be its first external financing. Alibaba’s Qwen 3.6 family has surpassed 700 million Hugging Face downloads — the most-downloaded AI model family in the world — under Apache 2.0. Moonshot’s Kimi K2.6 launched April 20 with frontier-class agent orchestration capability. Zhipu’s GLM-5.1 was trained entirely on a 100,000-chip Huawei Ascend cluster, the first fully Nvidia-independent frontier-class training run. The China sphere is operating at meaningfully lower absolute capability than the US frontier and meaningfully better cost-per-token economics. The capability gap is closing at roughly 6-12 months per cycle.
The European sphere I covered in the EU Sovereignty dispatch. Mistral at €11.7B valuation with €2.8B equity + $830M debt + 13,800 GB300 GPUs coming online in Q2. Aleph Alpha pivoted to PhariaAI orchestration platform with 70% compute reduction via T-Free architecture. Black Forest Labs leading on image and video generation with FLUX. The European sphere is structurally bounded — not competing for global frontier dominance but for EU regulated-market capture under the AI Act enforcement coming August 2.
The terrain has six Western labs, four major Chinese labs, and three European specialists in May 2026. Twenty months from now, the count is meaningfully different in every scenario.
II. The forces
Six forces are structuring the endgame. Each operates independently. The combinations of how each force resolves produce the scenarios.
Force 1 · Compute economics. The amortized cost of frontier training runs is growing 2.4× per year per Epoch AI’s measurements through 2024. GPT-4’s amortized cost was $40 million in March 2023; the trend implies $1 billion by early 2027 and $10 billion by 2028. Dario Amodei has stated publicly that single-run training costs will reach $10 billion in the next few years. The hardware acquisition cost is one to two orders of magnitude higher — GPT-4’s hardware acquisition cost was approximately $800 million against the $40 million amortized figure. By 2028, frontier training will require $10-50 billion of compute capital per training run.
The corollary: only labs with sustained access to that capital can maintain frontier competition. The capital is available today through three channels — sovereign wealth funds (concentrated in OpenAI and Anthropic), strategic investment from hyperscalers (Amazon-Anthropic, Microsoft-OpenAI, Nvidia and Salesforce-Reflection), and public markets (Anthropic IPO, OpenAI IPO probable 2027). Labs without one of these channels cannot reach frontier-2028 capability. This is the binding constraint that shapes Scenario A and the bounded version of Scenario B.
Force 2 · Capital availability and quality. The total private AI funding pool is large but concentrated. Q1 2026 saw $180 billion in AI funding, more than all of 2024. Roughly 80% of that capital flowed to OpenAI, Anthropic, and xAI. The remaining 20% spread across the spinout cohort, Reflection, and the rest of the industry. Sovereign wealth funds and private equity have absorbed nearly all of the marginal capital — OpenAI’s $50 billion Amazon commitment, $30 billion SoftBank, $30 billion Nvidia, plus the GIC, Coatue, Tiger anchor positions in Anthropic. The May 4 announcements that OpenAI and Anthropic each launched PE-backed enterprise joint ventures (Anthropic with Blackstone, Hellman & Friedman, Goldman Sachs at $1.5B; OpenAI’s Development Company at $10B with TPG, Brookfield, Advent, Bain) confirm that the most consequential capital relationships are now with alternative asset managers, not traditional VCs.
The corollary: capital is not the binding constraint for Anthropic and OpenAI. It is the binding constraint for everyone else. xAI’s capital comes through Musk’s personal-network channel, which is durable but unpredictable. Meta’s capital is internal and is paying a stock-price discount as a result. Google’s capital is internal and is paying a strategic-distraction tax (the Alphabet board cannot fund Gemini frontier investment at the same scale as Anthropic without broader corporate consequences). Reflection’s $2 billion is impressive for a Series A but is one or two cycles short of frontier-2028 capability. The spinout cohort is undercapitalized for frontier; they are positioned for vertical specialization.
Force 3 · Capability convergence and the open-weight floor. The Stanford AI Index 2026 reports that Chinese frontier models have “effectively closed” the gap to US frontier capability, trailing by 3-6 months on standard benchmarks. Open-weight models (Qwen, DeepSeek, Llama 4) are now within 6-12 months of frontier closed-weight capability. Alibaba’s Qwen 3.6 is the most-downloaded model family in the world. The DeepSeek V4 release on April 24 demonstrated that frontier-tier capability is now achievable on Huawei Ascend chips, which removes the Nvidia bottleneck for any state with sovereign chip access.
The corollary: frontier-tier capability is a depreciating asset on a 6-12 month cycle. Whatever Anthropic or OpenAI ship today, an open-weight equivalent will be available in 6-12 months at 1/20th the price per token. This force constrains how labs price, how labs disclose capability, and what the durable moat actually is. It is not the model. The model commoditizes. The moat is enterprise distribution, integration depth, regulatory-compliant deployment, and skills accumulation in the customer’s environment. This is what the forward-deployed engineer dispatch was about; it is what the skills marketplace dispatch was about; it is what the Channel Move dispatch was about. The capability convergence force shapes Scenario C heavily and undermines the assumed superiority that holds Scenario A together.
Force 4 · Talent flow. The 12 founders dispatch documented $3.4 billion in seed capital flowing to founders who departed Anthropic, DeepMind, OpenAI, Meta, and xAI between Q1 2025 and Q1 2026. The cohort is now bifurcating between the well-capitalized (Mira Murati’s Thinking Machines at $2B, Yann LeCun’s AMI Labs at $1.03B, David Silver’s Ineffable Intelligence at $1.1B) and the undercapitalized. xAI has lost all 11 of its original co-founders. DeepSeek is opening external financing largely to retain talent against Tencent, ByteDance, Xiaomi, and DeepRoute.ai poaching. Talent flow is now a leading indicator of which labs will reach frontier-2028.
The corollary: a lab losing its top researchers is a lab whose 2027 capability ceiling is being lowered today. The senior researchers who depart take with them deep knowledge of their former employer’s training methodology, capability profile, and weaknesses. They build companies that are explicitly competitive in the same capability space. The 2027-2028 frontier will not be competed for by the same six Western labs of 2026; it will be competed for by some of those labs plus 3-5 well-capitalized spinouts plus some that haven’t been founded yet.
Force 5 · Regulatory gating. The EU AI Act enforcement begins August 2, 2026, with €35M / 7%-of-global-revenue penalties. The Pentagon has split federal AI procurement into a multi-vendor classified channel (excluding Anthropic) and a single-source cybersecurity channel (Anthropic exclusively). The supply-chain risk designation against Anthropic is in active litigation. China has restricted domestic market access for foreign frontier providers since 2024. The U.S. is actively considering expanded export controls on frontier model weights and on chip access. AI is now a regulated industry across the major jurisdictions.
The corollary: regulatory positioning is now a primary variable in lab competitiveness. Anthropic’s strong EU and enterprise compliance posture maps well to the EU Act enforcement; its US federal positioning has been damaged by the supply-chain risk designation. OpenAI’s federal positioning is strong (Pentagon Channel 1 inclusion), but EU compliance overhead on closed-weight frontier models is significant. xAI inherits SpaceX’s federal positioning and is also constrained by Musk’s regulatory positions on speech. Mistral’s open-weight Apache 2.0 strategy gives it Article 53(2) procurement preference inside the EU. Each lab’s regulatory exposure is now part of its strategic position. This force interacts strongly with the tail-risk overlay.
Force 6 · The agentic transition. Q1 2026 was the quarter “agentic” stopped being a feature and became a category. Anthropic’s Claude Code revenue, OpenAI’s enterprise agent products, Google’s agent SDK, and xAI’s agent integration with X all matured to the point where agent-driven workloads are now meaningful percentages of revenue. The May 4 OpenAI/Anthropic enterprise venture announcements are explicit about the deployment model: forward-deployed engineers, Palantir-style integration, with PE-backed channel partners providing distribution. Agents are now the unit of economic value, not models.
The corollary: lab competitiveness through 2028 will be measured in agent-deployment revenue per customer, not in model-capability benchmarks. Labs with strong distribution (Anthropic via Channel 2 + Wall Street PE, OpenAI via Microsoft + Wall Street PE, Google via Cloud) can monetize even commodity-tier capability. Labs without strong distribution (Reflection, Meta to date, the spinouts) need to build it. This force is what makes Scenario C plausible — capability convergence does not eliminate the value, it just shifts the capture point.
These six forces operate together. The scenarios that follow are the three coherent ways they can resolve.
III. Scenario A: Two — the Duopoly Endstate
In this scenario, by Q4 2028, the Western frontier collapses to two scaled providers: Anthropic and OpenAI. Google DeepMind continues as an internal capability inside Alphabet, increasingly tied to Google Cloud distribution, but is not externally competing as a frontier-model-as-product company. Meta open-sources its Muse models and exits the frontier competition explicitly. xAI becomes a U.S. government-aligned vertical specialist tied to the SpaceX defense procurement channel, not a generalist frontier provider. Reflection AI is acquired by Microsoft, Google, or Nvidia by mid-2027. The spinout cohort produces 1-2 specialists that survive but none that reach frontier scale.
The forces resolving toward A. Compute economics resolve unfavorably for all labs without sustained $10B+ training run financing — meaning Anthropic, OpenAI, and Google. Capital availability concentrates further into the OpenAI/Anthropic-Plus-Google triad, with sovereign wealth funds deepening their positions and PE-backed enterprise channels capturing distribution. Capability convergence is real but slower than the optimistic case — Western open-weight stays 6-12 months behind frontier through 2028, which preserves the frontier-tier premium long enough for the duopoly to consolidate enterprise share. Talent flow stabilizes after the 2025-2026 spinout wave because the well-capitalized spinouts achieve specialty capability but cannot fund frontier training. Regulatory gating concentrates federal procurement around the largest providers (PE-backed channels effectively become the procurement mechanism). The agentic transition rewards distribution scale.
The end state. Anthropic and OpenAI have valuations of $1.5-2.5 trillion each. Both are public companies. Combined they capture 70-80% of enterprise frontier-AI spend in the US/allied sphere. Google DeepMind operates as a capability supplier into Google Cloud, with revenue counted under Google’s cloud line item — meaningful but not externally measurable as a frontier-model business. xAI becomes the defense AI vendor of choice for the U.S. government, with revenue concentrated in classified-environment workloads. Meta uses its open-weight Muse and Llama for advertising, content moderation, and AR glasses; it stops chasing frontier-tier capability. Reflection AI is acquired in 2027 for $15-25 billion as a strategic capability bolt-on. The spinout cohort either pivots to vertical applications, gets acquired, or fails.
The China sphere continues independently with 3-4 labs reaching frontier-tier capability (DeepSeek, Qwen, Moonshot, Zhipu) but operates in a structurally separate economic zone. The European sphere captures the EU regulated market with Mistral as the dominant player.
Probability estimate: 35%. This is the highest-probability scenario because it is the path of least resistance from where the forces stand in May 2026. The capital concentration is already at the level required for it to play out; the agentic-transition dynamics already reward the labs with PE-backed distribution channels; the spinout cohort, however well-capitalized, still has 18-30 months of training-and-product gap to close before they can compete on frontier deployment. The dominant risk to Scenario A is that the China sphere or open-weight cohort catches up to frontier capability faster than expected, which would compress the duopoly’s premium and force the structure into Scenario C.
Strategic implications. Capital allocation favors the duopoly. Enterprise CIOs simplify their procurement around two primary frontier vendors plus a regulatory-jurisdictional alternative (Mistral in EU, possibly DeepSeek-via-licensee in non-restricted markets). Investors should be long Anthropic + OpenAI, neutral on Google, short on Meta, neutral-to-short on xAI standalone (long via SpaceX), and selective on the spinout cohort. The risk to this allocation is the tail-risk overlay or scenario migration to C.
IV. Scenario B: Three-Plus-Sphere — the Equilibrium Endstate
In this scenario, by Q4 2028, three Western labs maintain frontier-tier scale: Anthropic, OpenAI, and Google DeepMind. xAI persists at sub-frontier-tier as a defense vertical. Meta operates as the open-weight provider, with Llama 5 / Muse 2 establishing meaningful enterprise share through cost-efficiency and open-source positioning. Reflection AI persists as a small frontier specialist, possibly via a strategic partnership with one of the three majors. 2-3 of the well-capitalized spinouts (Thinking Machines, AMI Labs, possibly Periodic Labs or Ineffable Intelligence) achieve near-frontier capability via specialization.
The China sphere stabilizes with 4 frontier-tier providers: DeepSeek, Alibaba (Qwen), Moonshot, and Zhipu. The European sphere has 2-3 viable players: Mistral (scale), Aleph Alpha (orchestration platform), and Black Forest Labs or Helsing (modality/defense specialists).
Total operating frontier-or-near-frontier providers globally in Q4 2028: approximately 10-12.
The forces resolving toward B. Compute economics resolve more favorably than in Scenario A because algorithmic efficiency improvements (lower-precision training, MoE architectures, better data efficiency) reduce the absolute training cost trajectory below the 2.4×/year curve. The 2028 frontier training run costs $5-8 billion rather than $10-15 billion, which keeps Google in the race and gives one or two well-capitalized spinouts the ability to reach frontier-tier with their existing capital. Capital availability stays robust for the top 4-5 Western labs as the Anthropic IPO succeeds, the OpenAI IPO follows in 2027, and sovereign wealth fund interest expands beyond the duopoly. Capability convergence is moderate — open-weight stays 9-12 months behind frontier closed-weight, preserving differentiation. Talent flow continues but stabilizes; the spinout cohort matures into specialist providers rather than frontier competitors. Regulatory gating produces a more multi-polar federal procurement environment — the Pentagon multi-vendor architecture (Channel 1 from the May 1 announcement) extends into civilian agencies. The agentic transition rewards both scale (Anthropic, OpenAI, Google) and specialization (Mistral, Reflection, the spinouts).
The end state. Anthropic, OpenAI, and Google DeepMind each have frontier-tier capability and meaningful enterprise distribution. Combined they capture approximately 55-65% of enterprise AI spend, with the remaining 35-45% distributed across xAI (defense), Meta (open-weight enterprise), Reflection (specialty), 2-3 spinout specialists, the European sphere (10-15% of EU regulated market), and the China sphere (in its own economic zone). Valuations: Anthropic $1.2-1.8T, OpenAI $1.5-2.0T, Google’s AI line $400-700B (notional), xAI $400-600B (post-SpaceX merger), Meta market cap implications neutral-to-positive, Reflection $40-80B (post-IPO or as private), Mistral €40-80B.
Probability estimate: 30%. This scenario is the second-most-likely because it requires a specific combination of moderately favorable resolution on multiple forces — algorithmic efficiency improvements that meaningfully reduce training costs, robust capital flow to the top 4-5 labs (not just top 2), sustained but not catastrophic capability convergence, and federal procurement that diversifies (which the May 1 Pentagon announcement is consistent with). This is the “everyone who is well-positioned today survives” outcome. It is plausible because the available data does not point unambiguously toward consolidation, and because the federal procurement architecture is already designed for multi-vendor.
Strategic implications. Capital allocation diversifies beyond the duopoly. Enterprise CIOs adopt 3-tier procurement: a top-tier frontier vendor, a cost-efficient near-frontier vendor (open-weight or Chinese-via-licensee), and a regulatory-jurisdictional alternative. Investors should be long Anthropic, OpenAI, and Google’s AI capability; positive on Mistral and Reflection; selective on the well-capitalized spinouts; and increasingly positive on Chinese frontier labs as US-China decoupling forces re-bundling of regional ecosystems. The risk to this allocation is the tail-risk overlay or scenario migration to A (consolidation) or C (stratification).
V. Scenario C: Stratification — the Tier Differentiation Endstate
In this scenario, by Q4 2028, the relevant question is no longer how many frontier labs survive; it is which tier of capability captures economic value. Anthropic and OpenAI are at the top of a “frontier-tier-1” capability range that approaches AGI-adjacent capability — agentic workflow execution, long-context reasoning that matches human expert performance, multi-modal grounded cognition. Google DeepMind has frontier-tier-1 capability but operates internally. Below the frontier-tier-1 cohort, a “frontier-tier-2” cohort of 6-12 providers (Meta, xAI, Reflection, Mistral, plus DeepSeek, Qwen, Moonshot, Zhipu, plus 2-3 spinouts) competes on commodity frontier capability — perhaps 6-12 months behind frontier-tier-1, at 1/10th to 1/20th the price per token. Below frontier-tier-2, a long tail of specialized open-weight derivatives competes on vertical specialization.
The economic question shifts. For 95% of enterprise workloads, frontier-tier-2 capability is sufficient. The top 5% — agentic workflows that genuinely require AGI-adjacent capability — pay frontier-tier-1 premium pricing. The bottom 80% — chatbots, document processing, code completion, content generation — runs on commodity-tier capability or open-weight derivatives. The model has commoditized; value moves to agents, skills, deployment, integration, and customer-specific judgment.
The forces resolving toward C. Compute economics resolve unfavorably for frontier-tier-1 sustainability — even Anthropic and OpenAI struggle with $10B+ training runs and the marginal capability improvements they buy are not commercially differentiated for 95% of workloads. Capability convergence is aggressive — Chinese open-weight closes to within 3-6 months of frontier; Western open-weight closes to within 6-9 months. The agentic transition is the dominant force, and it rewards distribution and integration depth far more than raw capability. Capital allocation shifts away from “more compute, bigger model” toward “more agents, better deployment.” The 12 spinouts cohort divides clearly between specialization successes (5-6 viable companies) and failures (the rest). Regulatory gating promotes interoperability and prevents single-vendor lock-in, which structurally favors the multi-tier outcome.
The end state. Frontier-tier-1: Anthropic, OpenAI, Google internal — the 3 labs that justify the AGI-adjacent capability premium. Frontier-tier-2: Meta, xAI, Reflection, Mistral, Aleph Alpha, plus 4 Chinese providers, plus 2-3 spinouts — 12 providers globally competing on commodity frontier capability. Open-weight derivatives: hundreds of specialized providers, much of the long tail concentrated around Llama 5, Qwen 4, and DeepSeek V5 base architectures. Total economic value capture: frontier-tier-1 captures 30-40% (premium pricing for the small number of workloads requiring AGI-adjacent capability), frontier-tier-2 captures 35-45%, and the long tail captures 20-30%. The frontier-tier-1 valuations are smaller than in Scenarios A and B because the addressable market is the 5% premium tier, not the full enterprise market.
Probability estimate: 25%. This scenario is the most demanding to forecast because it requires a specific resolution on the capability convergence force — open-weight has to close fast enough to commoditize the broad enterprise market while leaving a viable AGI-adjacent premium tier. Both halves of that have to be true. The data through May 2026 supports the first half (open-weight is closing quickly) but is ambiguous on the second (it is unclear whether AGI-adjacent capability is achievable on the 18-month horizon, and unclear whether enterprises will pay premium for it even if achievable). Scenario C is the highest-quality outcome for the broader economy because it produces the most competitive market, but it is not the most likely.
Strategic implications. Capital allocation requires careful tier-targeting. Investors holding only the frontier-tier-1 cohort capture less than they would in Scenarios A or B because the addressable premium market is smaller. Investors holding the frontier-tier-2 cohort (Mistral, Reflection, Meta open-weight, xAI vertical) capture meaningful share. Enterprise CIOs adopt explicit tier-mapping in their AI procurement — different workloads to different tiers, with explicit pricing-per-tier optimization. The agent layer (skills, integration, deployment infrastructure) becomes the highest-leverage investment opportunity, not the model layer. This is consistent with my Skills Marketplace dispatch — the marketplace position becomes the most defensible part of the post-model AI stack.
VI. The tail-risk overlay: crisis-triggered nationalization
A 15-25% probability event over the next 18 months is a Mythos-class capability proliferation incident that triggers crisis-mode regulatory response. The Mythos disclosure on April 7 demonstrated that frontier models can autonomously discover thousands of zero-day vulnerabilities. The disclosure was deliberately constrained — the consortium model, the SHA-3 commitments, the 99% unpatched ratio. The constraint depends on Anthropic and the Glasswing partners maintaining defensive control. Two leakage paths break the constraint.
Path 1. A Glasswing consortium member — there are approximately 40 organizations with access — has its access compromised. A nation-state actor or organized criminal group obtains Mythos-class capability access. The capability is used in a major cyberattack on critical infrastructure (financial system, power grid, healthcare network, telecommunications backbone). The political response is immediate and severe.
Path 2. Open-weight models reach Mythos-class offensive cybersecurity capability independently. Estimated timeline based on capability progression: 12-18 months from May 2026, putting it in the 2027 H1 to 2027 H2 window. Anthropic’s own framing of Mythos was that defensive deployment is racing against the clock for proliferation; the public statements of the consortium members make the same point. This is a forecast, not a hypothetical.
Either path triggers the same regulatory response. The U.S. government invokes Defense Production Act authorities to require frontier-capable models to operate under government oversight. A “Strategic AI Reserve” framework — modeled on the strategic petroleum reserve — provides government preferred-equity positions in Anthropic, OpenAI, and possibly Google’s AI capability. Sovereign-cloud deployment becomes mandatory for federal-classified workloads. The EU does similar via Article 7 (high-risk system) reclassification of all frontier models as Annex III high-risk, applying full conformity assessment requirements to model production, not just deployment. China closes its domestic market to foreign frontier providers and consolidates domestic capability under state direction.
The result. Whatever scenario was playing out is reshaped by the regulatory overlay. In Scenario A, the duopoly becomes regulated utilities with capped returns and state golden shares. In Scenario B, the federal procurement architecture compresses around 3-4 cleared providers with security clearances; the European and Chinese spheres tighten further; international model deployment is gated by export controls. In Scenario C, the stratification persists but the frontier-tier-1 cohort operates under government oversight, while the frontier-tier-2 and open-weight tiers face stricter export controls and possibly explicit capability ceilings.
The tail risk does not invalidate the base scenarios. It overlays them. Whichever scenario plays out, a Mythos-class proliferation event compresses returns, increases regulatory complexity, and shifts the equity structure of the major labs toward government-influenced governance.
Probability estimate: 15-25% in the 18-month window, rising to 30-40% in the 36-month window. The probability is not low. The capability proliferation that would trigger the response is on a trajectory that today’s data points to. The political-response architecture is not yet in place but is increasingly visible in legislative and executive deliberations. Tail-risk hedging is appropriate in any portfolio with significant frontier-AI exposure.
VII. The signposts
Fifteen leading indicators to watch over the 18-month forecast window. Each disambiguates between the scenarios in specific ways.
- Anthropic IPO pricing (October 2026). Pricing above $1 trillion implies Scenario A. Pricing $700B-$1T implies Scenario B. Pricing below $700B implies Scenario C or stress.
- OpenAI IPO timing announcement. Announcement before end of 2026 implies Scenario A or B. Delay to 2028 implies Scenario C or capital stress.
- Meta Q2 2026 capex revision. Pulled back to under $115B implies Meta accepts open-weight role (Scenarios B or C). Held or increased above $135B implies Meta still competing for frontier-tier-1 (Scenario B).
- Reflection AI productization. First commercial product launch with enterprise customers in 2026 implies Scenario B or C survival. No product by Q1 2027 implies acquisition (Scenario A).
- Microsoft strategic positioning vs OpenAI. Microsoft expanding own internal model development (Phi 5, etc.) implies Scenario B. Microsoft deepening OpenAI partnership exclusively implies Scenario A.
- Google DeepMind disclosures. Sustained $20B+ Q-over-Q cloud growth with explicit AI attribution implies Scenario B (Google viable as third frontier-tier-1). Cloud growth deceleration implies Scenario A.
- xAI capability disclosures vs SpaceX IPO. xAI hits frontier-tier benchmarks before SpaceX IPO implies Scenario B. xAI fails to hit benchmarks; SpaceX IPO is the real story implies Scenario A or sub-frontier specialization.
- DeepSeek V5 release timing and capability. V5 release within 9 months of V4 (i.e., by Q1 2027) at frontier parity implies aggressive capability convergence (Scenario C). Release delayed to mid-2027 or later implies Scenario A or B.
- Open-weight benchmark gap to frontier closed-weight. Gap closes to under 6 months by end of 2026 implies Scenario C. Gap holds at 9-12 months implies Scenario B. Gap widens implies Scenario A.
- Spinout cohort funding rounds (Thinking Machines, AMI Labs, Ineffable, Periodic). Frontier-tier valuations ($30B+) before end of 2026 imply Scenario B or C. Stalled or down-rounds imply Scenario A.
- Pentagon multi-vendor channel expansion. Channel 1 expansion to civilian federal agencies in 2026 H2 implies multi-vendor procurement persists (Scenarios B or C). Channel 1 consolidation to 2-3 vendors implies Scenario A.
- EU AI Act enforcement actions. First major non-compliance penalty against a U.S. hyperscaler within 12 months of August 2 implies regulatory architecture has teeth (relevant to all three scenarios but especially Scenario C). No penalties imply enforcement uneven (relevant to European bet failure mode 1).
- Sovereign wealth fund positioning. Continued capital concentration in OpenAI/Anthropic by GIC, Mubadala, ADIA, PIF, Norway implies Scenario A. Diversification to Google, xAI, Mistral, or spinouts implies Scenario B. Risk-off positioning implies stress.
- Mythos-class capability proliferation events. Any major incident or open-weight Mythos-class capability disclosure triggers tail-risk activation. No events by mid-2027 imply tail risk recedes but does not disappear.
- Talent flow direction. Net positive flow to the top three Western labs implies Scenario A. Net positive flow to spinouts and frontier-tier-2 cohort implies Scenarios B or C.
The signposts operate together. A pattern across multiple indicators is more meaningful than any single one. The first six months of enforcement (August 2026 – February 2027) should produce enough signal to identify which scenario is most consistent with the unfolding data.
VIII. What this means
The endgame question is not academic. It shapes capital allocation today.
For investors with public-equity AI exposure, the scenario probabilities I’ve assigned (A: 35%, B: 30%, C: 25%, plus 15-25% tail risk overlay) imply specific portfolio construction. A weighted-average outcome favors Anthropic and OpenAI heavily, but not exclusively. The expected value of a Reflection AI position is positive in B or C, neutral in A (pre-acquisition) or A-with-acquisition (small positive), and meaningfully positive in tail-risk scenarios where strategic capability acquisitions become more valuable. Mistral has positive expected value in B, C, and tail-risk; neutral in A. Meta has negative expected value in A, neutral in B, and positive in C. The implied portfolio is overweight Anthropic and OpenAI, neutral Google, modestly long Mistral and Reflection, modestly short Meta, and tail-risk hedged via instruments that benefit from regulatory tightening (defensive-AI vendors, sovereign-cloud providers, compliance platforms).
For private-market investors and venture capitalists, the spinout cohort is the highest-information asset class for the next 18 months. The 12 founders dispatch covered the cohort. The question now is which 3-5 of those 12 reach commercial scale. Watching the productization milestones and enterprise customer wins is more informative than watching the next funding round. The capital structure of the spinout cohort favors specialization-by-design — none of these companies has the capital to compete on generalist frontier-tier-1, so the question is which vertical or capability domain they own.
For enterprise CIOs and CTOs, the procurement architecture decision now is the architecture you live with through 2028. Single-vendor procurement around Anthropic or OpenAI maximizes integration depth at the cost of resilience to Scenarios B or C. Multi-vendor procurement maximizes resilience at the cost of integration overhead. The optimal middle ground for most enterprises is a primary frontier vendor (Anthropic or OpenAI), a secondary cost-efficient vendor (open-weight or Chinese-via-licensee for non-restricted workloads), and a regulatory-jurisdictional alternative (Mistral for EU, possibly xAI for federal). The skills marketplace dispatch covered why portable, cross-vendor skills are the hedge against vendor lock-in. That hedge is now operationally critical.
For policymakers, the tail-risk overlay is the priority. The base scenarios produce different but manageable industry structures; the tail-risk scenarios produce political crises that reshape the entire sector. Defensive AI capability for critical infrastructure, coordinated international response architectures for capability proliferation events, and clear governance frameworks for crisis-triggered Strategic AI Reserve mechanisms are the policy investments that pay off across all three base scenarios.
For Anthropic specifically, the IPO scheduling and pricing in October 2026 are the single most consequential decisions in the company’s history. A high-volume IPO at $1T+ valuation cements Scenario A. A measured IPO at $700-900B preserves optionality across B and C. A delayed or reduced IPO creates capital stress that could push the company into needing strategic acquisition by Microsoft, Google, or a sovereign entity. The IPO decision is being made under conditions where the strategic implications I’ve described are visible to the company’s leadership; the choice they make will be informative about how they view the scenario probabilities themselves.
IX. The endgame in one sentence
By the end of 2028, the Western frontier AI ecosystem will have either (A) consolidated to two scaled providers, (B) stabilized as a triad-plus-spheres of approximately ten globally active providers, or (C) stratified into a thin AGI-adjacent premium tier and a broad commodity-frontier tier — with a 15-25% probability that a Mythos-class capability proliferation event reshapes any of these outcomes through crisis-triggered nationalization.
The data through May 2026 does not yet identify which scenario is materializing. The next 18 months will. The signposts above are how to watch.
The capital, regulatory, and capability dynamics that will determine the outcome are mostly already visible. The question is not what will happen — it is which combination of forces will dominate. Forecasting is structured uncertainty, not narrative. The structure is the forecast. The numbers are the working hypothesis.
The endgame is six becoming two, three, or twelve. The bet you place today is the bet on which of those is real.
About the Author
Thorsten Meyer is a Munich-based futurist, post-labor economist, and recipient of OpenAI’s 10 Billion Token Award. He spent two decades managing €1B+ portfolios in enterprise ICT before deciding that writing about the transition was more useful than managing quarterly slides through it. More at ThorstenMeyerAI.com.
Related Pieces
- The Channel Move — Anthropic × Wall Street PE channel acquisition
- The 27% Problem — Anthropic’s enterprise lead and Google’s $750M check
- Single Digits — the April 2026 open-weight inflection
- The Bubble Is Not in Valuations — the productivity gap
- The Agent Trap — feature vs. infrastructure
Sources
- TechCrunch, Sources: Anthropic could raise a new $50B round at a valuation of $900B (2026-04-29)
- TechCrunch, Anthropic and OpenAI are both launching joint ventures for enterprise AI services (2026-05-04)
- Crunchbase News, Sector Snapshot: Venture Funding To Foundational AI Startups In Q1 Was Double All Of 2025 (2026-04-02)
- Tech-Insider, OpenAI Raises $122B at $852B Valuation (2026-04)
- CNBC, Meta debuts new AI model, attempting to catch Google, OpenAI after spending billions (2026-04-08)
- Fortune, DeepSeek unveils V4 model, with rock-bottom prices and close integration with Huawei’s chips (2026-04-24)
- Al Jazeera, China’s DeepSeek unveils latest models a year after upending global tech (2026-04-24)
- Renovate QR Research, Chinese AI Models in April 2026: DeepSeek V4, Kimi K2.6, Qwen 3.6 (2026-04)
- Stanford AI Index 2026
- USCC, Two Loops — How China’s Open AI Strategy Reinforces Its Industrial Dominance (2026-03)
- Epoch AI / Cottier et al., The Rising Costs of Training Frontier AI Models (arXiv 2405.21015)
- Dimension Research, Pretraining: The First Scaling Frontier (2025-12)
- Anthropic Frontier Red Team, Claude Mythos Preview (2026-04-07)
- ChinaTalk, DeepSeek V4 — The Post-DeepSeek Era (2026-04)
