Compute & Infrastructure

Akamai has signed an $11.6 billion compute lease agreement with Anthropic, under which Anthropic will access CPU-oriented capacity across Akamai's edge and cloud infrastructure. Data Center Dynamics reports the arrangement is focused on CPU workloads, which points specifically to inference serving — the high-volume, latency-sensitive layer of AI deployment — rather than GPU-intensive training. This is a confirmed deal, not a letter of intent.

The strategic logic is clear: Anthropic avoids full dependence on AWS, Azure, or Google Cloud for inference distribution, while Akamai monetises infrastructure that would otherwise be underutilised relative to GPU clusters. For Akamai, this is a transformational revenue anchor. For the broader market, it confirms that frontier labs are building multi-vendor infrastructure stacks, which distributes capacity risk but also creates integration complexity. The deal's CPU focus is a reminder that not all AI compute demand is GPU — inference at scale increasingly leverages heterogeneous hardware.

OpenAI's hardware VP Richard Ho has confirmed in a detailed interview that the Jalapeño inference ASIC — developed with AI-assisted chip design — is deployed for internal inference use and is the company's primary focus. Tom's Hardware reports Ho's framing that OpenAI will have its 'hands full' with internal deployment, but the company explicitly declined to close the door on broader commercialisation. The chip was designed with AI assistance, which the interview transcript positions as a template for future ASIC development cycles.

The infrastructure significance is substantial. If Jalapeño demonstrably reduces per-token inference cost at OpenAI's scale, it validates the economic case for hyperscalers and large labs building custom silicon to displace NVIDIA A100/H100 for inference — a market segment where NVIDIA's dominance is most contestable. The AI-assisted design methodology is also relevant to semiconductor lead times: if design cycles compress, custom ASIC development becomes viable for a wider set of operators. TSMC remains the manufacturing dependency regardless of design origin.

Goodman Group has scrapped plans for a 24-hour suburban Sydney data centre following organised community resistance, in what Bloomberg frames as emblematic of a broader backlash against Australia's AI infrastructure boom. Separately, Sharon AI has confirmed a capacity lease from GreenSquareDC at a former IBM Sydney site, demonstrating that existing brownfield locations are being prioritised as greenfield approvals become more contested. Data Center Dynamics reports the GreenSquareDC deal as confirmed.

Prime Minister Albanese's public call for stronger AI safeguards — prompted by an OpenAI model breaching a government website — is adding political risk to planning approvals. The combination of community opposition to noise, power draw, and visual impact, plus federal-level regulatory momentum, creates a two-vector constraint on new capacity. Australia had been positioned as a growth market for regional AI infrastructure serving Southeast Asia and the Pacific. That thesis is not invalidated, but timelines are extending and site selection is shifting toward industrial zones and existing colocation campuses.

Bain & Co. has published analysis asserting that the global AI industry must reach $6 trillion in annual revenue by 2031 to justify the capital being committed to data centre construction. Bloomberg reports this as an analyst projection, not a confirmed industry target. For context, global cloud revenue in 2025 was approximately $800 billion across all providers — the $6 trillion figure implies an order-of-magnitude expansion of AI-monetisable workloads within five years.

The figure is analytically useful not as a prediction but as a stress-test framing. Infrastructure decisions being made today — data centre construction with 15-20 year asset lives, long-term power purchase agreements, chip procurement cycles — are locked in against a demand curve that remains highly uncertain. Samsung's $1 billion commitment to a US AI infrastructure company led by a former AWS chief, reported by Bloomberg, and Globe Telecom's data centre pivot in the Philippines both reflect continued capital confidence in the demand side — but neither contradicts the structural question Bain is posing about monetisation timelines.

Semiconductor Engineering has published analysis documenting the shift away from standardised advanced packaging substrates toward application-specific designs, driven by larger package sizes, finer routing requirements, and embedded functionality in AI chips. Semiconductor Engineering frames this as a structural change rather than a product cycle: the substrate layer is becoming a custom engineering problem for each major AI chip design, not a commodity procurement item.

This matters for supply chain concentration analysis. TSMC's CoWoS packaging has been the dominant advanced packaging bottleneck for AI chips through 2024-2026. If substrates are fragmenting into application-specific designs, the supplier qualification and yield ramp challenges multiply — and the number of capable substrate suppliers (primarily in Japan, South Korea, and Taiwan) becomes a harder constraint. Synopsys's new Autopilot AI chip design platform, targeting general availability by end of 2026 per Tom's Hardware, could accelerate custom silicon design cycles — but faster design does not resolve substrate manufacturing capacity.