Token pricing remains radically uncertain despite supply crunch

South Korea commits $919B to sovereign AI infrastructure; Motif Technologies trains best non-Chinese open-source model on $15M budget

South Korea announced a massive $919B investment across chip manufacturing, AI robotics, and data centers, targeting 18.4GW of capacity by 2035 (8.4GW by 2029) led by SK Group, GS Group, and Naver. The announcement repackages several existing projects but signals serious state commitment to AI independence following the country's "Independent AI Foundation Model" tournament. The tournament results reveal a striking efficiency signal: Motif Technologies, a sub-30-person startup with just $17M raised and access to 768 B200s, trained Motif-3 on approximately $15M in compute costs and achieved the highest score on Artificial Analysis' Intelligence Index (66 points), surpassing established competitors backed by major chaebols. Motif-3 ranks ahead of both Inkling and Nemotron 3 Ultra—the best American open-source models—despite being less than half their size. However, the Korean government eliminated Motif in the latest round despite its benchmark dominance, citing opaque expert review and user testing criteria, raising questions about the tournament's evaluation methodology.

Anthropic releases Claude Fable 5.1 and Mythos 5.1 with 75% cache-read price cut; tops Intelligence Index at 66 points

Anthropic released Fable 5.1 and Mythos 5.1, achieving a score of 66 on Artificial Analysis' Intelligence Index—the highest measured score, ahead of prior leader Opus 5 (63) and GPT-5.6 Sol (61). The models show significant improvements on coding and agentic benchmarks, with Fable 5.1 posting 91.4% on Terminal-Bench v2.1 and 62.0% on SciCode, representing 2x+ gains over Fable 5 on some coding tasks. Pricing remains flat on per-token input/output ($10/$50 per million tokens), but Anthropic cut cache-read costs 75% from $1.00 to $0.25 per million cached tokens. Despite the cache savings, Fable 5.1 costs approximately 20% more per task than Fable 5 because output token usage increased ~1.7x, resulting in a net $3.76 per Intelligence Index task versus $3.13 for Fable 5. Both models include invisible watermarking for EU AI Act compliance and Mythos 5.1 includes reduced safeguards for cybersecurity and life sciences work. The release signals Anthropic's confidence in iterative improvements and aggressive cost positioning ahead of OpenAI's Astra release.

Nvidia reportedly acquiring Hugging Face for $12.9B to control model orchestration layer

Nvidia has agreed to acquire Hugging Face, the open-source model repository, for $12.9 billion according to The Information. The deal reflects Nvidia's strategy of using its massive balance sheet ($127B annual FCF) to acquire control points across the AI stack, particularly middleware and orchestration layers that increase customer lock-in and reduce supplier bargaining power. This mirrors Stripe's $7.5B acquisition of OpenRouter, both betting on LLM commoditization requiring dominant middleware platforms.

Vercel's AI SDK software factory authors 25-35% of merged PRs and closes 70-80% of issues within four weeks

Vercel deployed an AI-powered software factory for its AI SDK (20M weekly npm downloads, 26K GitHub stars) that autonomously processes issues and pull requests using specialized agents for triage, bug reproduction, fixing, and review. Four weeks after deployment, the factory authors 25-35% of merged PRs and closes 70-80% of issues, demonstrating concrete productivity gains that justify replacing traditional community contributions. The factory uses dedicated agents for specific tasks (bug reproduction, fixes, PR review, backports, documentation, feature analysis, feature implementation) rather than a single multi-skilled agent, enabling easier maintenance and debugging. All agents run in isolated Vercel Sandboxes with network shielding to prevent exfiltration of secrets or malicious code execution, with human merge gates on all changes. The system processes GitHub webhooks automatically and includes a monitoring UI for reviewers. This represents a concrete signal of how AI agents are beginning to replace human labor in high-leverage infrastructure maintenance tasks.

Nvidia NVSwitch emerging as dominant fabric for AI scale-up networks, analogous to InfiniBand's historical role

NVSwitch is becoming the de facto standard interconnect for large-scale AI clusters, reinforcing Nvidia's control over the full AI compute stack beyond GPUs. Nvidia accounts for approximately 95% of GPU revenues and 75% of combined GPU/XPU revenues, giving it dominant share of scale-up networking through NVLink/NVSwitch. The company is not selling NVSwitch ASICs separately—customers must purchase Nvidia "Grace" or "Vera" CPUs or "Blackwell"/"Rubin" GPUs, and Nvidia-supplied racks with NVSwitch fabric. Nvidia generated approximately $25.74B in trailing twelve-month revenue, with roughly equal contributions from Ethernet and NVSwitch interconnects combined. However, competing standards like UALink (backed by AMD, Broadcom, Cisco, Google, HPE, Intel, Meta, Microsoft) and ESUN/SUE-T are advancing, with UALink 1.0 capable of scaling to 1,024 devices versus NVSwitch's 576-device limit across two tiers.

SB Energy files for IPO with 8.8GW contracted capacity and $430B data center backlog; heavily dependent on OpenAI

SB Energy, the SoftBank-backed AI power infrastructure company, filed for IPO targeting $5-7B in proceeds, disclosing 8.8GW of contracted capacity and a $430B data center backlog. The company is "substantially dependent" on OpenAI as both a tenant and equity investor, with OpenAI CEO Sam Altman as an early personal investor. None of SB Energy's data centers are operational yet; the company incurred $3.2B in net losses in H1 2026 while generating $139M in revenue from legacy energy business. Nvidia announced $105B in financing for an OpenAI data center in Ohio to be built by SB Energy, with CEO Rich Hossfeld citing Nvidia's involvement as critical for unlocking investment-grade financing. The IPO prospectus flagged growing public backlash against AI infrastructure as a risk factor, noting more than half of Americans express greater concern than excitement about AI adoption.

Elon Musk planning to manufacture gas turbine blades and build 10GW capacity by next year amid severe power supply constraints

Current market leaders (GE Vernova, Siemens, Mitsubishi) have ~50GW annual capacity sold out through 2030, creating a critical bottleneck for data center buildout; Musk's vertical integration mirrors his broader strategy of controlling the entire AI infrastructure stack from chips (SpaceX foundry) to power generation to satellite data centers.

Ways to think about token pricing

Benedict Evans in Ben Evans published an essay titled "Ways to think about token pricing". Main takeaways:

• Token pricing today reflects a supply crunch, but all variables (supply, demand, marginal cost, ROI) remain radically uncertain over 3-5 year horizons, making confident forecasts unreliable.
• The core question is whether foundation models capture sustainable pricing power or become low-margin commodity infrastructure with value captured upstream (like cellular networks or TSMC).
• Four key uncertainties will determine outcomes: frontier model demand elasticity across use cases, continued frontier capability improvements, competitive dynamics among frontier labs, and whether value concentrates in models or in tooling/products built atop them.
• Mobile data provides a more apt comparison than fiber: massive capacity buildout, marginal cost economics, enormous industry scale ($1T revenue), but stocks went nowhere and value was captured elsewhere.
• Current inference margins (40-50% gross) exclude training costs (currently far larger than revenue) and assume unknown asset lives; profitability depends on revenue scaling faster than training cost growth.

Building a Software Factory for AI SDK

Vercel in Vercel Blog published an essay titled "Building a Software Factory for AI SDK". Main takeaways:

• Factory agents now author 25-35% of weekly merged PRs and close 70-80% of issues, reducing peak open issues from 1,022 to 844 in 6 weeks
• Architecture uses specialized single-task agents (bug reproduction, fixes, review, backports, docs) rather than multi-skilled agents, improving maintainability and debuggability
• Human reviewers remain in control; factory generates comprehensive risk assessments (side-effect, performance, backwards-compatibility) to calibrate review depth by change risk
• Security implemented via isolated Vercel Sandboxes with network shielding and secret containment; all merges require human approval
• Concrete example: issue #17898 (blocked-domain support) moved from request → classification → analysis → implementation → review → merge → backports in fully documented agent workflow

South Korea Announces $919B Investment into Chip Manufacturing, AI Robotics, and Data Centers; Plans 18.4GW Data Center Capacity by 2035

DataCenterDynamics reported on South Korea's sovereign AI infrastructure investment announcement. Main takeaways:

• South Korea plans $919B total investment across semiconductors, AI robotics, and data centers to 'win the AI race,' with $352B committed to first phase (2028-2029)
• Data center buildout targets 8.4GW by 2029 and 18.4GW by 2035, led by SK (5GW), GS (2.4GW), and Naver (1GW) in phase one; would multiply current capacity sevenfold
• Samsung and SK Hynix commit $513B for four memory fabs (two each) in Jeonnam-Gwangju plus $52B for HBM packaging, signaling major supply-side consolidation
• All projects scheduled to break ground H1 2028 with phased operations from 2029, representing coordinated state-backed infrastructure buildout competing directly with US and China

In The Long Run, Nvidia NVSwitch Is The InfiniBand Of Scale Up AI Networks

The Next Platform published a strategic analysis of Nvidia's NVSwitch interconnect technology and its competitive position against emerging Ethernet-based alternatives. Main takeaways:

• Nvidia currently dominates scale-up networking with ~95% GPU market share and ~75% combined GPU/XPU share, but this advantage is temporary as hyperscalers prefer open standards
• UALink (backed by AMD, Broadcom, Google, Meta, Microsoft, and others) can theoretically scale to 1,024 devices vs. NVSwitch's 576, with potential for InfiniBand-like latencies (~100ns) atop Ethernet infrastructure
• NVLink Fusion (allowing third-party chips to integrate NVLink ports) is Nvidia's defensive play to stall UALink/ESUN adoption, but historical precedent (InfiniBand's decline despite technical superiority) suggests Ethernet standards eventually win through scalability and interoperability
• Nvidia's networking revenue split shows 19.6% of datacenter sales in Q2 F2027, double typical cloud ratios, reflecting rackscale NUMA architecture requirements
• Even if UALink becomes dominant, Nvidia will likely capture significant share by building best-in-class UALink switches, similar to its InfiniBand dominance trajectory