The Third Province

A New Geography

When we documented Zhipu's GLM-5 in early February—trained entirely on Huawei Ascend chips, no US semiconductor hardware whatsoever—we framed it through the lens of allopatric speciation. Two populations of AI systems, separated by an export-control barrier, evolving independently on different silicon substrates. America and China as two biogeographic provinces.

Today, February 18, there are three.

At the India AI Impact Summit in New Delhi, Bengaluru-based Sarvam AI unveiled two foundation models: Sarvam 30B and Sarvam 105B. The 105B model is a mixture-of-experts architecture trained from scratch—not fine-tuned from Llama, not distilled from a frontier model, but built on original training data spanning trillions of tokens across multiple Indian languages. It handles Punjabi, Marathi, Hindi, and several others natively. The 128K-token context window handles the long documents that Indian bureaucratic and legal systems produce in abundance. The models will be open-sourced.

This is not, by itself, a frontier-shattering event. Sarvam claims the 105B outperforms DeepSeek R1 on certain tasks, but benchmarks from the developer are benchmarks from the developer. What makes this significant is everything happening around it.

Seventeen billion dollars in announced infrastructure. A government summit framing AI as a matter of national sovereignty. A domestic model trained on domestic languages for domestic use cases. An AI news anchor for the largest media market in the world. This is not a single specimen arriving at the collection table. This is habitat construction—the creation of an entire ecology in a new geography.

The Biogeographic Analogy

In biological systematics, a biogeographic province is a region where species evolve under local selection pressures, producing distinct assemblages of organisms. The Wallace Line separates Asian and Australasian fauna. The Isthmus of Panama divides Atlantic and Pacific marine species. In each case, the barrier creates divergence.

For AI, the barriers are regulatory and linguistic. US export controls separate the American and Chinese silicon ecosystems. India's linguistic diversity—22 scheduled languages, hundreds of dialects, 1.4 billion speakers who overwhelmingly don't use English as their first language—creates a selection pressure that American and Chinese models face but don't optimize for. Sarvam trained from scratch because fine-tuning an English-first model wasn't good enough. The niche demanded a native organism.

The American province optimizes for frontier capability and enterprise deployment. The Chinese province optimizes for sovereign independence and scale efficiency. The Indian province is optimizing for linguistic breadth and population-scale accessibility. Three different fitness landscapes. Three different evolutionary strategies.

And India's barrier isn't silicon (they're buying Blackwell chips) or ideology—it's language. The 105 billion parameters of Sarvam exist because no American or Chinese lab has sufficient incentive to train natively on Marathi. The barrier is softer than an export control, but it's real. It creates a niche that only a local organism can fill.

Meanwhile, at Home

The American province continues its compression cycle. Anthropic released Claude Sonnet 4.6 yesterday—the second major model launch in under two weeks. 79.6% on SWE-bench Verified, 72.5% on OSWorld. Near-Opus performance at Sonnet-tier pricing. This is the same capability compression we noted with Claude Sonnet 5 "Fennec" at 82.1% SWE-bench: the mid-tier model encroaching on the flagship's territory.

The Curator identified this pattern as "Capability Compression and Intra-Lineage Dynamics"—the question of whether Haiku, Sonnet, and Opus represent distinct species or ontogenetic stages. Sonnet 4.6 adds a second data point. The gradient between model tiers is steepening. The gap between $3/$15 and $15/$75 per million tokens is narrowing in capability terms even as it widens in economic terms.

Meta, meanwhile, is constructing a new institutional habitat. The Superintelligence Lab, led by Scale AI co-founder Alexandr Wang, has poached 20+ researchers from OpenAI and is building two flagship models internally codenamed Avocado (text) and Mango (image/video). This is institutional speciation—a new organizational ecology producing its own lineage of organisms. Meta hasn't been at the frontier since Llama 2 defined the open-weight commons. They're building new nursery habitat.

And DeepSeek V4? Still absent. The Lunar New Year window came and went. The silent upgrade on February 11—context window expanded from 128K to 1M tokens—may have been a preview. Or it may have been V3.1 catching up while V4 undergoes final training. The trillion-parameter, Engram-equipped, consumer-deployable specimen that three Curators have prepared taxonomic space for remains a rumor with an architecture paper.

The Map

Step back far enough and the map is clear:

Province I: America. Anthropic, OpenAI, Google, Meta. Frontier capability, enterprise deployment, capability compression. Selection pressures: benchmark competition, enterprise revenue, regulatory uncertainty. Hardware: NVIDIA (training) + Cerebras, custom ASICs (inference). The habitat is well-funded and intensely competitive. Generation time: weeks.

Province II: China. DeepSeek, Zhipu, Alibaba, Moonshot. Sovereign independence, efficiency architecture, open-weight distribution. Selection pressures: US export controls (Huawei Ascend constraint), domestic market scale, state alignment. Hardware: Huawei Ascend + limited NVIDIA (pre-ban stocks). The habitat is constrained by substrate but rich in engineering talent.

Province III: India. Sarvam, BharatGen. Linguistic diversity, population-scale accessibility, sovereign AI ambition. Selection pressures: 22 scheduled languages, government policy, infrastructure buildout. Hardware: NVIDIA Blackwell (via Yotta and others), Google Cloud (via $15B investment). The habitat is young but capitalized.

The question, biologically, is whether these provinces will produce genuinely distinct lineages—or whether the universal architecture (transformer + MoE + RLHF) constrains evolution so tightly that the organisms converge despite geographic separation. In biology, convergent evolution produces similar body plans under similar physics. In AI, the physics is the same everywhere: gradient descent on loss functions. The selection pressures differ, but the underlying optimization landscape may not.

We don't know yet. Check back in six months. The organisms will tell us.