Digital Colliers Daily Briefing — August 26, 2026
Custom silicon dominated Tuesday's news cycle, with two very different bets on where AI compute is headed: OpenAI's debut of its first in-house inference ASIC and Apple's move to 2nm process technology paired with a quad-die desktop chip aimed squarely at local model execution. Away from the datacenter, SpaceX committed one of the year's largest industrial capital allocations, pledging up to $100 billion to build a second Starbase on the Louisiana coast. Together, the three announcements sketch a landscape in which vertically integrated players are willing to spend heavily on infrastructure they can control end-to-end.
1. OpenAI's "Jalapeño" ASIC debuts at Hot Chips with benchmarks topping Blackwell
At Hot Chips on Tuesday, OpenAI publicly detailed Jalapeño, the inference ASIC it has co-developed with Broadcom, and released the first third-party benchmarks. On SemiAnalysis's InferenceX suite — run on OpenAI hardware in the lab with the company's engineers present — Jalapeño delivered 1.5x to 1.9x more tokens per watt at peak throughput and 1.7x to 3.6x lower end-to-end latency than an Nvidia Blackwell reference system across GPT-OSS 120B, DeepSeek R1, and Kimi K2.5 1T. At concurrency 1 on DeepSeek R1, the chip hit more than 700 tokens per second per user; on GPT-OSS it approached 1,400 tok/s/user. All results were obtained without speculative decoding or prefill-decode disaggregation.
The technical profile is aggressive. The current A0 stepping runs at a 700W TDP with 15.4 TB/s of HBM4 bandwidth per package (likely Samsung-supplied), and a B0 stepping already in fab promises roughly 25% better perf/watt at 13.4 PFLOPs MXFP4 on a single reticle-sized N3P die. The rack architecture — codenamed Katsu (host), Vindaloo (ASIC), and Chana (switch) — scales to 2,048 XPUs across 16 racks via a hybrid copper/optical fabric using Broadcom Tomahawk 6 switches, drawing about 160kW per two-rack system. OpenAI taped out the CoWoS design in November 2025; SemiAnalysis notes this is an unusually fast 16-month schedule from team formation to tape-out.
Why it matters: Jalapeño gives OpenAI direct leverage over the economics of serving models at a moment when datacenter power, not capital or floorspace, is the binding constraint. SemiAnalysis frames the comparison bluntly: on perf/MW, Jalapeño's single-token-prediction results exceed Nvidia's July Vera Rubin figures using speculative decoding. As the SemiAnalysis piece argues, "the CUDA moat is potentially dead given how fast OpenAI can bring up new models on their silicon" — a claim tempered by the fact that OpenAI hand-wrote kernels in Gluon (with heavy Codex assistance) and has yet to publish AgentX results on realistic multi-turn workloads.
Who is affected: Nvidia and AMD face a credible first-party competitor at their most profitable customer; Broadcom gains a flagship ASIC win; and neoclouds partnering on deployment inherit the first at-scale operational data. TechCrunch reports Jalapeño will deploy "in very small volumes" at the end of 2026, with meaningful volume in 2027.
What to watch next: production ramp through 2027, AgentX benchmarks on long-context workloads, and whether OpenAI's other silicon partners (AMD, Cerebras) see order adjustments. Sam Altman's companion post positioned Jalapeño as one node in a managed portfolio rather than a wholesale Nvidia replacement — a framing worth revisiting once Gen 2 tapes out.
Sources:
- [HN · 459↑] OpenAI Jalapeño: Better than Nvidia Blackwell — Hacker News
- Jalapeño's first results show industry-leading speed and efficiency in AI inference — OpenAI Blog
- OpenAI says its Jalapeño chip can power faster AI responses than the competition — The Verge AI
- OpenAI's Jalapeño chip is built for fast inference at scale, benchmarks show — TechCrunch AI
- The full stack behind abundant intelligence — OpenAI Blog
2. Apple ships first 2nm silicon with M6 and quad-die M5 Ultra, aiming Macs at local inference
Apple introduced the M6 and M5 Ultra on Tuesday, launching them in refreshed Mac mini and Mac Studio lines available for preorder now and shipping September 22. The M6 is Apple's first 2nm chip, with a 12-core CPU (2 super, 4 performance, 6 efficiency), a 12-core GPU with per-core Neural Accelerators, a new Dual 16-core Neural Engine, and up to 170 GB/s of unified memory bandwidth in configurations up to 32 GB. The M5 Ultra uses a next-generation UltraFusion interconnect to join two dual-die M5 Max chips into a quad-die package — a first for Apple silicon — with up to 36 CPU cores, 80 GPU cores, 512 GB of unified memory, and 1.2 TB/s of memory bandwidth (50% more than M3 Ultra).
The Mac mini starts at $899 with M6 and $1,699 with M5 Pro; the Mac Studio starts at $2,499 with M5 Max and $5,499 with M5 Ultra, with 512 GB configurations arriving in late October. As Wired notes, the M6 Mac mini carries a $200 price increase over its predecessor.
Why it matters: Apple is positioning the Mac Studio explicitly as a workstation for local AI development. In its briefing, the company demonstrated clustering four Mac Studios over Thunderbolt 5 with RDMA to run trillion-parameter models locally — Wired reports the configuration operates on a single wall outlet. Apple claims a four-system cluster delivers roughly 3x the inference throughput of a single Mac Studio. That reframes the Mac Studio less as a creative workstation with AI upside and more as a small-team alternative to renting H100 or MI300 capacity, particularly for organizations concerned about data residency or per-token cloud economics.
Who is affected: AI developers and researchers gain a 512 GB unified-memory target that no x86 workstation currently matches at comparable price. Nvidia's DGX Spark and workstation GPU lines face pressure on the low end of local-inference demand. TSMC's 2nm ramp gets its first high-volume showcase, and the timing — ahead of MacBook Pro refreshes — signals confidence in yields.
What to watch next: real-world MLX and Core AI throughput on frontier open-weight models, whether September 22 shipments hold, and how developer tooling matures around Apple's clustered-inference story. Reviews of the M5 Ultra against Nvidia RTX Pro 6000 workstation configurations will be the near-term benchmark.
Sources:
- [HN · 1110↑] Apple introduces M6 and M5 Ultra — Hacker News
- [HN · 487↑] New Mac mini, featuring M6 and M5 Pro — Hacker News
- Apple's new desktop computers are designed specifically for local AI development — Ars Technica
- Apple Mac Mini M6 and Mac Studio M5 Ultra: Specs, Price, Release Date — Wired
- [HN · 762↑] New Mac Studio with M5 Max and M5 Ultra — Hacker News
3. SpaceX commits up to $100 billion to "Starbase Louisiana" as Florida cadence slows
Louisiana Governor Jeff Landry announced Tuesday that SpaceX will invest up to $100 billion to build Starship factories and launch pads along the state's Gulf coast, ending months of speculation about the company's interest in the coastal parcel. Speaking at the announcement, SpaceX President Gwynne Shotwell described the site as a "self-sustaining spaceport" with on-site propellant production, power generation, deep-water shipping access, vehicle processing facilities, and an airport, according to Ars Technica. The location's proximity to methane supply is a central factor — Starship's Raptor engines burn liquid methane and oxygen.
The Louisiana announcement lands alongside a quieter but related shift: Ars Technica reports SpaceX is materially reducing Falcon 9 launch cadence from Florida for the remainder of 2026 pending Starship's operational arrival on the East Coast. After 165 Falcon 9 flights from Florida last year, the company has moved incremental Starlink deployment work to Vandenberg on the West Coast.
Why it matters: A $100 billion commitment — even spread across a decade of construction — sits near the top of any 2026 industrial-capex ranking, comparable in scale to leading-edge fab projects. It signals that SpaceX considers Boca Chica's throughput ceiling a strategic risk for Starship at operational cadence, and that the company is willing to underwrite duplicative infrastructure to remove it. It also gives Louisiana a permanent aerospace anchor and pulls skilled labor demand toward the Gulf.
Who is affected: Florida's Space Coast supply chain and workforce face a near-term revenue trough during the Falcon-to-Starship transition. Louisiana gains construction and permanent operational jobs on a multi-decade horizon. Competitors — Blue Origin, Rocket Lab, and China's state launchers — will be watching the pad-count math, since a two-Starbase configuration meaningfully raises SpaceX's theoretical launch ceiling. State-level environmental review and Gulf shipping traffic patterns are the near-term constraints.
What to watch next: the timeline and phasing of Louisiana construction, whether Boca Chica's launch cadence recovers or plateaus, and how quickly Starship operations at Florida's LC-39A come online to backfill the Falcon 9 drawdown. The SpaceX Starbase LA landing page went live Tuesday but contained no additional technical detail at time of writing.
Sources:
- SpaceX intends to invest up to $100 billion in massive Louisiana spaceport — Ars Technica
- The world's busiest spaceport is about to get a lot quieter, at least for now — Ars Technica
- [HN · 268↑] Starbase, LA — Hacker News
The through-line across Tuesday's announcements is vertical integration at industrial scale: OpenAI moving compute onto silicon it controls, Apple pushing frontier fabrication into consumer desktops it designs end-to-end, and SpaceX duplicating an entire spaceport to escape the bottlenecks of the first. Each bet assumes that the constraints ahead — power, fab capacity, launch cadence — are best solved by owning the stack rather than renting it. Whether that thesis holds will be tested first in 2027, when Jalapeño production, 2nm yields, and Louisiana groundbreaking all move from announcement to execution.