Firmus Partners with Nvidia to Build $30 Billion AI Data Center in Indonesia

Firmus Technologies, an Australian AI infrastructure company, has announced an eight-year partnership with Nvidia to build a 360-megawatt "Nvidia DSX AI Factory" campus on Batam Island, Indonesia. The campus will be developed jointly with Singapore-based DayOne, with offtake agreements expected to carry commitments of up to $30 billion over six years starting from the first year.

Batam Island sits just off the coast of Singapore, making it a strategically important data center location. As competition over AI infrastructure accelerates across Southeast Asia, the fact that Firmus has secured large-scale power capacity through the Nvidia partnership is drawing attention. With AI demand surging, the race to build data centers in the Asia-Pacific region looks set to intensify even further.

Token Costs for AI Coding Reach Parity with Developer Salaries

According to CIO.com, token costs in AI-assisted software development are reaching levels comparable to traditional developer labor costs. As the use of AI coding tools becomes routine, the cumulative cost of API calls is creating a heavier financial burden than anticipated.

This gives companies a critical decision point: whether the simple equation of "boosting development efficiency with AI" holds up now depends on cost management. For enterprises deploying AI assistants at scale, optimizing token consumption has become a new technical challenge. In short, we have entered a phase where companies must carefully evaluate whether API costs offset—or even exceed—the labor savings.

CEO Nadella: "Every Company Should Have Its Own AI Model"

Microsoft CEO Satya Nadella has argued that every company should build its own AI models. His message: reducing dependence on general-purpose models and owning models that leverage proprietary data and domain knowledge will become a source of competitive advantage.

The remarks emphasize AI democratization and corporate autonomy, but realizing this vision involves technical hurdles and cost challenges. While it is unrealistic for every company to train models from scratch, the statement can be interpreted as pointing toward building "company-specific" AI systems through fine-tuning existing models or combining them with RAG (retrieval-augmented generation).

DeepSeek Open-Sources Speculative Decoding Framework "DeepSpec"; DFlash Merged into llama.cpp

DeepSeek has open-sourced "DeepSpec," a speculative decoding framework that speeds up LLM inference, developed as joint research with Peking University. It includes three draft model algorithms—DSPark, DFlash, and Eagle3—and also provides trained checkpoints for the Qwen3 series and Gemma-4-12B.

Separately, the LocalLLaMA community on Reddit reports that DFlash support has been merged into llama.cpp, making speculative decoding standard in local environments. DFlash is said to have improved generation speed by up to 85% in DeepSeek-V4 production while maintaining throughput, and it is drawing attention as a technology that could significantly raise the practicality of local LLMs.

Anthropic Freezes Claude Agent SDK Billing Change on the Day of Enforcement

Anthropic froze, on the very day it was to take effect, a billing change that would have excluded the Claude Agent SDK from subscription coverage. For developers running Claude Code in CI/CD pipelines or automated reviews via GitHub Actions, a sudden cost increase would have directly hit production workflows, so the decision has been widely welcomed.

Among developers, unglamorous but practical workflows have taken root—such as using `claude -p` in overnight CI jobs to summarize test failure logs and post them to Slack. Reconsidering the billing policy for these "not constantly supervised by a human" use cases marks a critical juncture in the adoption phase of agentic AI.

Robotaxi Rollout: Reality Slower Than Expected

A Bloomberg feature analyzed how Waymo, Tesla, Baidu, and others are operating robotaxi businesses, yet adoption is not accelerating as fast as expected. Since Google announced its self-driving car project in 2010, improvements in safety and efficiency have been touted, but regulatory barriers and logistical challenges still stand in the way.

Over the past year, robotaxi services have launched in Croatia, Singapore, Dubai, Abu Dhabi, and Riyadh, and London is expected to follow shortly. Still, robotaxis have yet to replace conventional taxis at scale. It is becoming clear that even technically feasible technology takes longer than expected to achieve real-world deployment.

Concerns Over Economic Risks Posed by the AI Boom

The Sydney Morning Herald highlighted the risks that overheated AI investment ("AI exuberance") poses to the global economy. The piece argues that massive AI investment could trigger a recession or the erosion of the middle class, drawing attention as a warning about the economic side effects of the AI boom.

If AI investment pays off in the form of productivity gains, there is no problem—but if returns fall short of expectations, negative effects on the broader market are a concern. Striking the right balance between technological progress and economic stability looks set to become a key medium-term challenge going forward.

---