Modal Raised $750M at a $15.75B Valuation After Its Revenue Tripled to $1B ARR in Five Months. Serverless GPU Just Became a Decacorn Category.
On October 15, 2026, Microsoft announced it finished pre-training MAI-Base-2, an 80-billion-active-parameter mixture-of-experts model with 2.8 trillion total parameters. The company claims 4–12x better compute efficiency than DeepSeek-V4-Pro-Base and, according to reports, CEO Mustafa Suleyman told staff it beats OpenAI's GPT-5.6 at one-quarter the compute. The model is still in post-training and hasn't shipped — but the architecture and efficiency claims signal Microsoft's most significant bet on AI independence from OpenAI.
Microsoft announced on October 15, 2026, that it has completed the pre-training phase of MAI-Base-2, its internally developed frontier AI model. The model is a mixture-of-experts (MoE) architecture with 2.8 trillion total parameters and 80 billion active parameters per token — comparable in scale to the largest publicly disclosed frontier models, but with a significantly different efficiency profile. Microsoft claims compute efficiency 4 to 12 times better than DeepSeek-V4-Pro-Base on its held-out evaluations, and according to reporting from 24/7 Wall St., CEO Mustafa Suleyman told Microsoft staff the model beats OpenAI's GPT-5.6 at approximately one-quarter the compute.
MAI-Base-2 has not shipped. The model completed pre-training and is now in post-training — the alignment, instruction tuning, and safety evaluation cycle that typically adds months to a deployment timeline. Microsoft plans to license the weights through Foundry, its enterprise AI platform, to trusted partners rather than releasing them publicly. But the completion of pre-training is itself a significant announcement, independent of any product timeline.
This is the first time Microsoft has disclosed a frontier-scale pre-training run of this magnitude built entirely on its own research stack. It is a direct signal of Microsoft's intent to establish AI independence from its OpenAI investment — and it has implications for enterprise buyers, AI infrastructure investors, and every team currently building on OpenAI APIs.
What MAI-Base-2 Is: Architecture and Scale
MAI-Base-2 is a mixture-of-experts model, a class of neural network architecture that has become the dominant design for frontier-scale systems since GPT-4's reported MoE structure and DeepSeek's open disclosure of its MoE training approach. The core principle of MoE is parameter efficiency: rather than activating every parameter for every token, a routing mechanism selects a subset of "expert" sub-networks to process each token. A model can therefore have far more total parameters than a dense architecture without proportional inference cost, because only a fraction of those parameters operate at any given moment.
For MAI-Base-2, the scale breaks down as follows:
- Total parameters: 2.8 trillion — the full set of weights stored in the model
- Active parameters per token: 80 billion — the expert sub-networks used to process each token
- Active-to-total ratio: approximately 2.9%, consistent with other large MoE frontier models
- Training data: approximately 50 trillion tokens, with 18% vision-language data (images paired with descriptive text)
- Architecture properties: natively multimodal, capable of processing both text and images without separate vision encoders
The 18% vision-language training mix is notable. Most frontier models treat multimodal capability as an add-on — a vision encoder bolted onto a text-dominant base model. Training 18% of total tokens on vision-language data from the start suggests Microsoft's intention to build genuine multimodal capability into the base rather than layering it on during post-training.
The Training Run: 38 Days on 16,000 GB200 GPUs
Microsoft's official pre-training announcement describes a 38-day training run on 16,000 NVIDIA GB200 GPUs — the current generation of Blackwell architecture compute, combining a B200 GPU with a Grace CPU and NVLink Switch chip in a rack-scale unit designed for large-scale AI training. Note that some secondary reports cite 6,000 GPUs, likely conflating a subset configuration with the full deployment; Microsoft's materials cite 16,000.
A 38-day run on 16,000 GB200 GPUs represents compute expenditure in the range of the largest publicly disclosed frontier training runs. The choice of NVIDIA GB200 hardware rather than Microsoft's own Maia 200 custom AI chip — which is serving production Copilot workloads as of October 2026 — is an important detail: Microsoft trained MAI-Base-2 on third-party hardware rather than its custom silicon, suggesting that Maia 200's production-training capability is not yet at the scale required for a 2.8-trillion-parameter run.
The data mix for the training run includes 50 trillion tokens at approximately 18% vision-language data. For comparison, DeepSeek's V3 technical report disclosed approximately 14.8 trillion tokens of training data. MAI-Base-2's training data footprint is roughly 3.4x larger than DeepSeek V3's — a comparison that helps contextualize the efficiency claims, since a larger training corpus allows the model to achieve a given capability level at smaller active parameter count.
The Efficiency Claims: What 4–12x Better Than DeepSeek Actually Means
Microsoft's claim of 4 to 12 times better compute efficiency over DeepSeek-V4-Pro-Base is the most technically significant and most carefully qualified element of the announcement. The claim is based on Microsoft's own held-out evaluations — not independently verified benchmarks — and the range reflects variation across benchmark categories rather than a single number.
The practical meaning of "4–12x better compute efficiency" depends on the comparison frame:
Token-for-token inference cost. If MAI-Base-2 matches DeepSeek-V4-Pro-Base's output quality on a task while consuming 4 to 12 times fewer floating-point operations per token, the inference cost per response would be correspondingly lower. At enterprise inference scale — hundreds of millions of queries per month — a 4x inference cost reduction is material budget relief.
Quality-normalized efficiency. The range likely reflects that different task types favor different architectures. On tasks where DeepSeek-V4's MoE routing is well-calibrated, the efficiency advantage may be at the low end of the range. On tasks that benefit from MAI-Base-2's larger training corpus or architectural improvements, the gap widens.
Training compute efficiency. The claims may also imply that MAI-Base-2 reached a given capability threshold with less training compute relative to earlier Microsoft pre-training runs. This is harder to verify without disclosure of comparative training runs from Microsoft's earlier models.
| Metric | MAI-Base-2 | DeepSeek-V4-Pro-Base | Efficiency comparison |
|---|---|---|---|
| Total parameters | 2.8T | ~685B (reported) | ~4.1x larger |
| Active parameters / token | 80B | ~37B (reported) | ~2.2x larger |
| Training tokens | ~50T | ~14.8T (disclosed) | ~3.4x more data |
| Claimed compute efficiency | 4–12x advantage | Baseline | Microsoft's own benchmarks |
| Independent verification | Pending Foundry partners | Open-weights, publicly tested | Not yet comparable |
Sources: Microsoft announcement, DeepSeek technical report. All figures approximate; independent verification pending.
The Suleyman Comparison: Beats GPT-5.6 at One-Quarter the Compute
The most striking element of the MAI-Base-2 announcement is Mustafa Suleyman's reported internal statement that MAI-Base-2 beats OpenAI's GPT-5.6 using approximately one-quarter of the compute. The comparison is notable on multiple dimensions.
First, the fact that it was made at all — Microsoft's own AI chief, in internal communications, benchmarking their new model favorably against OpenAI's best — would have been inconceivable two years ago when Microsoft's AI strategy was defined by the OpenAI partnership and Suleyman himself joined from DeepMind to accelerate Microsoft's first-party AI capabilities. The comparison signals a cultural reorientation at Microsoft AI from "how do we best deploy OpenAI's models" to "how does our own model compare."
Second, "one-quarter the compute" aligns mathematically with the 4x end of the 4–12x efficiency range Microsoft claims against DeepSeek. If MAI-Base-2 is 4x more efficient than DeepSeek-V4 on a compute-normalized basis, and GPT-5.6 is also roughly comparable to DeepSeek-V4 in efficiency, the transitive math supports the one-quarter-compute claim — though this chain of comparisons requires independent verification to treat as specification.
Third, the comparison was internal rather than public, which limits its accountability. Independent evaluators will test these claims when Foundry partners get access. Until then, enterprise teams should treat the GPT-5.6 comparison as a directional signal from Microsoft's internal benchmarking — credible enough to warrant attention, not verified enough to be a specification.
Microsoft's AI Independence Strategy
Google's Gemini 4 Argon benchmark lead was framed as a capability milestone. OpenAI's DevDay GPT-6.1 Sol was framed as an enterprise cost reduction. Microsoft's MAI-Base-2 announcement is framed differently from either: it is a statement of capability, not a product announcement, and its strategic significance runs deeper than any benchmark result.
Microsoft has invested reportedly more than $13 billion in OpenAI across multiple rounds since 2019. That investment secured exclusive cloud computing arrangements, early model access, and the right to deploy OpenAI models through Azure. The commercial logic was sound: Microsoft gained AI product capability it could not have built from zero in the timeframe that mattered, and OpenAI gained the Azure compute and commercial distribution it needed to grow. Both companies benefited.
But every strategic investment in a third-party capability carries a structural risk: if the third party's capability advantage erodes, the investor's competitive position erodes with it. As IBM demonstrated with its self-hosted, on-premises approach to enterprise AI, and as Anthropic demonstrated with its $100M enterprise training investment, the 2026 enterprise AI market is bifurcating between platform owners and platform tenants. MAI-Base-2 is Microsoft's clearest signal that it intends to be a platform owner.
The Foundry licensing model — distributing weights to trusted partners rather than releasing them publicly or keeping them exclusively API-accessible — represents a strategic middle path. OpenAI keeps its weights proprietary and distributes only through API. Meta releases weights openly under limited commercial licenses. Microsoft is positioning for controlled commercial distribution: closed weights, selective partner access, enterprise commercial relationships. This preserves control over the model while allowing Microsoft to seed enterprise ecosystems without the open-release risks of weight availability to bad actors.
The Post-Training Phase: What Happens Between Pre-Training and Availability
Pre-training completion is a significant technical milestone, but not the same as a model being ready to use. Post-training for a frontier-scale model of MAI-Base-2's size and commercial stakes typically involves multiple distinct stages, each adding weeks to months:
Supervised fine-tuning (SFT). Training the base model on curated datasets of human-written demonstrations — instruction-response pairs, multi-turn conversations, task completions — to teach the model to follow instructions and respond in ways that are helpful and coherent for user-facing applications.
Reinforcement learning from human feedback (RLHF). Training a separate reward model based on human preference comparisons between model outputs, then using reinforcement learning to optimize the base model's responses toward those preferences. This phase is computationally expensive and requires significant human labeling infrastructure — both the computational cost and the human review time add to the timeline.
Safety evaluation and red-teaming. Systematic testing for harmful outputs, jailbreaks, prompt injection attacks, and misuse vectors. For a model at MAI-Base-2's capability level, safety evaluation is not a checkbox exercise. Microsoft has an enterprise AI safety team, and the evaluation program for a frontier model of this scale runs for months, not weeks.
Enterprise-specific fine-tuning. Depending on the Foundry use cases, post-training may include vertical fine-tuning for code, legal, medical, or financial applications — tailoring the base model to the capability profiles that enterprise customers prioritize.
A model that completed pre-training in mid-October 2026 and moves through a thorough post-training process could reach Foundry trusted partners in Q1 2027 — optimistically. More conservatively, H1 2027. Enterprise teams should not build Q1 2027 plans around MAI-Base-2 availability.
What MAI-Base-2 Means for Enterprise AI Buyers
The announcement has four practical implications for enterprise teams currently evaluating or deploying frontier AI models.
The vendor landscape is diversifying faster than procurement timelines can track. Enterprise AI model procurement in 2024 was largely a choice between OpenAI via Azure, Anthropic via AWS Bedrock, and Google Vertex AI. By Q4 2026, the landscape includes IBM's self-hosted Bob, Google Gemini 4 Argon, OpenAI GPT-6.1 Sol, Nous Research's Hermes, and now a pending Microsoft-native frontier model. Procurement teams that locked in multi-year agreements in 2024 should review whether their contracts include model flexibility clauses.
Microsoft Azure enterprise agreements may soon include first-party frontier model access. If MAI-Base-2 performs as claimed in independent evaluation, it could become the default model powering Copilot and Azure AI services — replacing the OpenAI dependency for workloads where performance is comparable or better. Enterprise teams on Azure should monitor their AI model agreements for changes and build relationship with their Microsoft account team now to understand the roadmap.
The efficiency claims, if verified, have material cost implications. Enterprise AI inference is now a significant operating expense for large organizations. A model delivering comparable performance at 4–12x lower compute cost would reduce that cost materially at the scale most enterprises operate. This is not a claim to act on yet — it is a claim to test once Foundry access is available.
The Foundry licensing model creates a new procurement option for regulated industries. Enterprise teams that need on-premises or dedicated-cloud access to a frontier-quality model without dependence on a public API — financial services, healthcare, defense — should register for Foundry partner status now to position for early access to a model that Microsoft controls and can deploy in configurations that meet data residency requirements.
Four Steps for Enterprise Teams Following MAI-Base-2
1. Audit your current OpenAI API dependency through Azure. Map which production workloads run on Azure OpenAI services, the monthly inference volume each generates, and the cost structure per workload. This audit creates the baseline for comparing any Microsoft-native frontier model as an alternative — and the evidence base for negotiating model flexibility into any Azure AI contract renewals in 2027.
2. Register for Foundry trusted partner status. Microsoft's Foundry partner program is the path to early access to MAI-Base-2 before general availability. Enterprise AI teams not already in the Foundry ecosystem should begin the partner registration process now. The evaluation period between pre-training completion and GA is the window to establish the commercial relationship and qualify for early access.
3. Cap frontier model contract commitments at twelve months. The frontier model landscape is moving faster than multi-year contract structures can accommodate. Any enterprise agreement signed in Q4 2026 that locks in model terms for three years risks being on the wrong model when MAI-Base-2, Gemini 5, or Claude 5 reaches general availability. Negotiate annual review or model portability terms into all new AI infrastructure contracts.
4. Build model-flexible AI architecture from the start. The teams that will benefit most from MAI-Base-2 are those whose AI applications were built with model-swap capability built in from the beginning — abstraction layers that allow switching the underlying model without rewriting application logic. For teams that have not yet built this abstraction, the MAI-Base-2 announcement is the latest in a series of reminders that the frontier model landscape requires architectural flexibility, not deep vendor coupling.
The Platform Independence Bet
The 2.8 trillion-parameter model is not primarily a product. It is a demonstration of capability and a strategic signal. Microsoft is informing the market that it can train frontier-scale models on its own research stack, at efficiency levels competitive with the best open and closed models in the world. The company is not announcing that MAI-Base-2 will replace OpenAI's models in Copilot or Azure AI tomorrow. It is establishing that, if it needed to, it could.
That capability changes Microsoft's position in every subsequent conversation about the OpenAI partnership, about Azure AI pricing, and about enterprise model vendor strategy. A company with no frontier-capable in-house model has limited negotiating leverage with its AI model suppliers. A company that has demonstrated it can train a model that outperforms GPT-5.6 at one-quarter the compute — in its own CEO's internal assessment — has substantially more leverage, regardless of what MAI-Base-2's eventual independent benchmarks show.
For enterprise buyers, the strategic takeaway is structural: the frontier AI market in Q4 2026 is genuinely competitive. OpenAI, Google, Anthropic, and now Microsoft all have credible frontier capabilities. That competition puts downward pressure on model pricing, increases the importance of vendor flexibility, and reduces the risk that any single provider achieves lock-in through capability monopoly. The teams best positioned to benefit from this competition are those who are not locked in already.
Takeaway: Microsoft completing MAI-Base-2 pre-training is the strongest signal yet of a genuinely competitive frontier model market in 2026. The efficiency claims — 4–12x better than DeepSeek-V4-Pro-Base, one-quarter the compute of GPT-5.6 by internal benchmarks — require independent verification before enterprise teams treat them as specifications. But the strategic reality requires no verification: Microsoft now has a first-party frontier model it owns, controls, and can deploy independently of the OpenAI relationship. Enterprise teams building on Azure should begin the Foundry partner registration process, audit their OpenAI model dependencies, and hold AI infrastructure contracts to annual commitment terms. The model landscape that exists in Q1 2027 will be materially different from the one that existed in Q1 2026 — MAI-Base-2 is one of the reasons why.
Frequently Asked Questions
What is Microsoft MAI-Base-2?
MAI-Base-2 is a frontier-scale AI foundation model developed internally by Microsoft AI. The 'MAI' prefix stands for Microsoft AI, and the '2' designation indicates it is the second generation of Microsoft's frontier pre-training effort. Microsoft announced the completion of the model's pre-training phase on October 15, 2026. MAI-Base-2 is a mixture-of-experts (MoE) architecture with 2.8 trillion total parameters and 80 billion parameters active per token during inference. The model was trained on approximately 50 trillion tokens, with 18% of the training data being vision-language data (images paired with text), making it natively multimodal. MAI-Base-2 is not yet available to users — it is currently in the post-training phase, which includes supervised fine-tuning, reinforcement learning from human feedback, and safety evaluations. Microsoft plans to distribute the model weights to trusted partners through its Foundry platform rather than releasing them publicly, following a model similar to the selective licensing approach used by other frontier AI companies.
What are MAI-Base-2's architecture and parameter count?
MAI-Base-2 uses a mixture-of-experts (MoE) architecture, the dominant design for frontier-scale AI systems in 2025–2026. MoE models divide the parameter space into a set of 'expert' sub-networks, with a routing mechanism that selects a subset of experts to process each token during inference. This design allows the model to have a large total parameter count — in MAI-Base-2's case, 2.8 trillion — while activating only a fraction of those parameters per token, which reduces inference compute relative to a dense model of the same total parameter size. MAI-Base-2 activates 80 billion parameters per token, approximately 2.9% of its total 2.8 trillion parameters — an active-to-total ratio consistent with other large MoE frontier models. The architecture also incorporates vision-language capabilities natively, trained on image-text pairs representing 18% of the total 50 trillion token training corpus. Microsoft trained the model over 38 days on 16,000 NVIDIA GB200 GPUs, the current generation of Blackwell architecture hardware designed for large-scale AI training.
What efficiency claims is Microsoft making about MAI-Base-2?
Microsoft claims MAI-Base-2 achieves 4 to 12 times better compute efficiency than DeepSeek-V4-Pro-Base on the company's held-out evaluation benchmarks. The range — 4x to 12x rather than a single figure — reflects variation across different benchmark categories: the efficiency advantage is larger on some task types and smaller on others. 'Compute efficiency' in this context means the model can match DeepSeek-V4-Pro-Base's output quality on those benchmarks while requiring 4 to 12 times less compute to generate a given response. At inference scale, this translates to lower cost per query for the same capability level. Microsoft also claims, based on internal benchmarks, that the model outperforms every open-weight model the company tested. These are Microsoft's own held-out evaluations, not independently verified by third-party benchmarks — the distinction matters for enterprise teams evaluating the claims, who should treat them as directional signals until Foundry partners run independent evaluations on production workloads.
When will Microsoft MAI-Base-2 be available?
MAI-Base-2 is not yet available as of October 2026. The model completed pre-training on October 15 and is currently in the post-training phase, which for a frontier-scale model typically involves multiple stages: supervised fine-tuning on curated instruction-following demonstrations, reinforcement learning from human feedback (RLHF) to align the model with user preferences, direct preference optimization (DPO) as an alternative or complement to RLHF, and extensive safety evaluation including red-teaming for harmful outputs, jailbreaks, and misuse vectors. For a model of MAI-Base-2's scale and the evident commercial and safety stakes of the deployment, the post-training phase typically runs three to six months. A model that completed pre-training in mid-October 2026 could reach Foundry's trusted partner network in Q1 or Q2 2027, assuming post-training evaluation does not surface issues that require additional remediation. Microsoft has not announced a specific availability timeline. Enterprise teams interested in early access should begin the Foundry trusted partner registration process now rather than waiting for a general availability announcement.
What does MAI-Base-2 mean for Microsoft's relationship with OpenAI?
MAI-Base-2's existence signals a meaningful strategic evolution in Microsoft's AI posture relative to OpenAI, though it does not represent a rupture in the partnership. Microsoft has invested reportedly more than $13 billion in OpenAI across multiple rounds since 2019, securing exclusive cloud computing arrangements, early model access, and the right to deploy OpenAI models through Azure. That investment remains intact, and OpenAI models continue to power Copilot and Azure AI services. But MAI-Base-2 establishes that Microsoft no longer needs OpenAI's models to train and deploy a frontier-capable AI system. The hedge logic is straightforward: any enterprise strategy that is strategically dependent on a supplier relationship it does not control carries structural risk. If OpenAI's models underperform Google's Gemini or Anthropic's Claude in future benchmarks, Microsoft's enterprise AI competitiveness would be exposed without an in-house alternative. MAI-Base-2 is that alternative. The Foundry licensing model — weights to trusted partners, not open release — mirrors Microsoft's preference for controlled commercial distribution over open-source availability.
How should enterprise teams respond to the MAI-Base-2 announcement?
Enterprise teams should treat the MAI-Base-2 announcement as a planning signal rather than a current product decision. The model is not yet available, the efficiency claims are Microsoft's own and unverified by independent benchmarks, and the Foundry distribution model limits initial access to trusted partners rather than general availability. The near-term actions that make sense are: register for Foundry partner status to position for early access; audit current OpenAI model usage through Azure to create a baseline for future comparison; hold frontier model contracts to annual commitment terms rather than multi-year agreements that would lock in model terms before MAI-Base-2 evaluation is possible; and wait for independent benchmark results from Foundry partners before making procurement decisions based on Microsoft's efficiency claims. The medium-term implication is more structural: the frontier model market is becoming more competitive, not less. Enterprise teams that build flexible AI infrastructure — able to swap models as the quality and cost landscape evolves — will outperform teams that locked in specific model dependencies before the competitive landscape settled. MAI-Base-2 is one more data point in support of model-flexible architecture.