Mythos: Sovereign AI Scale Race, Memory Crunch

In March 2026, Anthropic described Claude Mythos Preview as its most capable model to date and a step change in performance. Fortune first reported the exposed pre-release material, after which Anthropic acknowledged that training was complete and limited testing was underway. The model's debut pulled frontier competition back toward large-scale pretraining.

Anthropic did not disclose the model's size, but industry estimates place Mythos in the 10T-class range. Its capability leap strengthened the outlook that a larger pretraining base can create a wider performance gap. DeepSeek then moved to raise $7.4 billion, while xAI formally announced that it was training a 10T model. The center of frontier competition had shifted back toward hyperscale.

Mythos Revived the Large-Scale Pretraining Race

At the starting point of that hyperscale race was Mythos's capability leap. Anthropic's Mythos Preview system card called it the company's most capable frontier model at the time and described a striking leap across multiple evaluations versus Claude Opus 4.6. The AISI evaluation likewise found stronger autonomous vulnerability discovery and multistage attack performance than the previous generation.

The market cared less about an exact parameter count than the size of the performance jump. The official evaluation results and the unofficial estimate of a much larger pretraining base revived the view that scale could again produce a major gap. Mythos brought the parameter race back to the center of frontier AI.

DeepSeek Raises $7.4B as xAI Trains 10T Model

The shock quickly became a change in investment strategy. According to The Information, DeepSeek—fresh from leading the price competition through cost efficiency—saw the Mythos preview and concluded that frontier competition required more capital. Liang Wenfeng moved to raise $7.4 billion. DeepSeek's move showed that cost efficiency alone was not enough to close the frontier gap; competing at that level still required capital for larger-scale training.

xAI/SpaceXAI is currently training a 10T-labeled model on Colossus 2. In an official X post on April 8, 2026, Elon Musk said it was among seven models in training, alongside two 1T variants, two 1.5T variants, one 6T model, and Imagine V2. Musk used the 10T label without defining whether it referred to total or active parameters. DeepSeek's capital expansion and xAI's current run both point back toward hyperscale competition.

GPT-6 Scale Rumors, Gemini's New Pretraining Base

Behind the moves by DeepSeek and xAI came more scaling signals from OpenAI and Google. The AI Daily Brief, citing leakers Leo and Andrew Curran, reported the rumor that GPT-6 will use a new pretraining base with substantially more parameters. Epoch AI predicts that GPT-6 will return to a larger pretraining run using more compute than GPT-4.5. The parameter scale remains a rumor, but the compute forecast adds weight to the outlook for a larger GPT-6.

Gemini 3.5 Pro provides another signal in the same direction. Business Insider reported that the planned June launch slipped to July, while TechTimes reported that Google abandoned the prior base for a new pretraining cycle and checkpoint. The delay and base-reset report together suggest that Google favored a fresh pretraining foundation over further work on the existing checkpoint.

The GPT-6 rumor, Epoch AI forecast, and Gemini reports come from different sources and carry different levels of certainty. After Mythos raised the competitive bar, however, all three pointed toward the same hyperscale direction: larger models, more training compute, and new base checkpoints.

Sovereign AI Intensifies HBM and DRAM Demand

The same hyperscale direction is spreading beyond a few US and Chinese clusters to the national level. Governments that want cyber, defense, administrative, and health data to remain inside national borders are currently expanding sovereign AI infrastructure. The Center for a New American Security Sovereign AI Index catalogs operational and planned national AI factories and sovereign clouds across Europe, the Middle East, and Asia.

When each jurisdiction builds its own training and inference capacity, AI data-center demand spreads beyond a single global cluster. Persistent agents, autonomous vehicles, and robotics add server DRAM and storage demand alongside accelerators.

HBM is a central bottleneck for large-model training. Larger models and checkpoints raise capacity and bandwidth requirements per accelerator, while production diverted toward HBM can tighten general-purpose DRAM supply. The HBM demand outlook from SK hynix's newsroom puts the 2026 market at $54.6 billion, up 58% year over year. Pressure on HBM and DRAM supply is rising sharply.

Samsung Electronics, SK hynix, Micron, and other core memory suppliers are increasingly positioned as the main beneficiaries of a prolonged memory supercycle. Large-scale pretraining competition and the spread of sovereign AI data centers are lifting HBM and DRAM demand together.