Digital Colliers Daily Briefing — July 16, 2026
Three structurally significant stories anchor today's briefing, each pointing at a different pressure point in the AI and platform economy. Thinking Machines Lab has finally shipped its debut model, reframing the open-weights conversation around a U.S.-based lab. TSMC's raised 2026 guidance offers the cleanest read yet on whether AI capex is still climbing. And a collapsed Epic settlement is about to force Google to distribute competing app stores from inside Google Play — a rare instance of an antitrust remedy reshaping platform mechanics rather than merely fining the platform.
1. Thinking Machines ships Inkling, its first open-weights model, and reframes the U.S. open-source stack
What happened. Thinking Machines Lab, the Mira Murati–led startup that raised the largest seed round on record at a $12 billion valuation, released Inkling, a 975B-parameter Mixture-of-Experts model with 41B active parameters, a 1M-token context window on the open weights, and native multimodal inputs across text, image, and audio. The company also previewed Inkling-Small, a 276B/12B-active sibling. Weights are on Hugging Face under Apache 2.0, and the model launched with day-0 support on vLLM, SGLang, Modal, Baseten, Databricks, TogetherAI, Fireworks, and Hugging Face Transformers, plus an NVFP4 checkpoint for Blackwell inference. The company trained Inkling from scratch on 45 trillion multimodal tokens using Nvidia GB300 NVL72 systems, according to its release post.
Why it matters. Thinking Machines is explicit that Inkling is "not the strongest overall model available today, open or closed." Instead, per the company's own framing, it is pitched as a customizable base — a substrate for Tinker, the fine-tuning platform that is TML's actual revenue mechanism. Artificial Analysis places Inkling at 41 on its Intelligence Index, ahead of Nvidia's Nemotron 3 Ultra (38), Gemma 4 31B (29), and gpt-oss-120b (24), making it the leading U.S.-based open-weights release — though still trailing GLM-5.2, Kimi K2.6, and DeepSeek v4 Pro on several axes, as Latent Space summarizes. As TechCrunch notes, the release lands amid a broader argument — echoed this week by Satya Nadella and Hugging Face's Clem Delangue — that enterprises using closed frontier models effectively pay twice, once in fees and again in leaked domain expertise.
Who is affected. Enterprises building on private data, particularly those uncomfortable with API-based capture of prompts and corrections; the Chinese open-weights labs (Moonshot, DeepSeek, Zhipu) that have dominated the open leaderboard; Nvidia, which is embedded across TML's training and inference stack; and the fine-tuning-as-a-service segment, which now has a credible American MoE base with unusually broad ecosystem support. Architectural choices — relative positional encoding instead of RoPE, 5:1 sliding-window-to-global attention, short convolutions around attention and FFN branches — are drawing attention as potential leading indicators for future frontier designs.
What to watch next. The full release of Inkling-Small; whether Tinker fine-tunes translate into paying enterprise deployments; independent verification of the Bridgewater result (84.7% on financial reasoning at ~1/14th proprietary cost, per TechCrunch, though only self-reported); and whether the next TML model can be trained without the Kimi K2.5 SFT bootstrap, as the company has said it intends. Kimi K3 and DeepSeek V4 GA are expected imminently and will reset the comparison.
Sources:
- [HN · 965↑] Inkling: Our Open-Weights Model — Hacker News
- Thinking Machines amps up its bet against one-size-fits-all AI with its first open model, Inkling — TechCrunch AI
- [AINews] Thinky's Inkling: 975B-A41B multimodal, new best American Apache 2.0 open model (with Inkling-Small, 276B-A12B) — Latent Space
- Thinking Machines Lab Drops Its First Model — Wired
- not much happened today — smol.ai News
- INSANE AI News: GPT-RED, Kimi K3, Gemini 3.5 Pro and Anthropic's "END GAME" — YouTube · Wes Roth
2. TSMC's raised 2026 guidance closes the debate on near-term AI capex
What happened. Taiwan Semiconductor Manufacturing Co. raised its 2026 capital expenditure range from $52–56 billion to $60–64 billion and lifted USD revenue growth guidance from "30%+" to "40%+" year-over-year, explicitly citing the "AI megatrend," according to Bloomberg's Debby Wu. The upward revision accompanied Q2 results: revenue of roughly $39.45 billion, up 36% YoY, and net income of about $21.9 billion, up 77.4% — both above consensus. Chips at 7nm and below accounted for 77% of wafer revenue, per CNBC.
Why it matters. As the sole leading-edge foundry for Nvidia, AMD, Apple, and every serious AI accelerator design, TSMC's guidance functions as the industry's most credible demand indicator. A mid-year upward revision of $8 billion in capex and a ten-point revenue-growth bump suggests hyperscaler and merchant-silicon orders for 2026 are firming, not softening — a pointed counterweight to the recurring "AI capex plateau" thesis. The 77% concentration in 7nm-and-below also confirms that the mix continues to shift toward the nodes where AI accelerators, high-end mobile SoCs, and advanced packaging compete for the same limited CoWoS capacity.
Who is affected. Nvidia most directly, given its dependence on TSMC N3/N4 wafer starts and CoWoS-L packaging — the GB300 NVL72 systems TML used to train Inkling are downstream of exactly this capacity. AMD's MI-series roadmap and Apple's silicon cadence share the same bottleneck. Equipment suppliers — ASML, Applied Materials, Lam Research, Tokyo Electron — see the raised capex as direct forward orders. Hyperscalers building custom silicon (Google TPU, AWS Trainium, Microsoft Maia) compete for the same slots. On the sovereign side, the guidance reinforces Taiwan's centrality to AI supply and complicates any near-term diversification narrative around U.S. and Japanese fab buildouts.
What to watch next. Whether the capex uplift translates disproportionately into CoWoS advanced-packaging capacity, which remains the tightest constraint; TSMC's commentary on N2 ramp timing at its next earnings call; and any signals about Arizona and Kumamoto output as a share of leading-edge wafer starts.
Sources:
- TSMC raises its 2026 capex projections from $52B-$56B to $60B-$64B, and ups revenue growth projection from 30%+ to 40%+ YoY in USD terms, citing "AI megatrend" (Debby Wu/Bloomberg) — Techmeme
- TSMC reports Q2 revenue up 36% YoY to ~$39.45B, net income up 77.4% YoY to ~$21.9B, both above est., and says chips 7nm or smaller were 77% of its wafer revenue (Jenny Lee/CNBC) — Techmeme
3. Google Play forced to host rival app stores as Epic settlement collapses
What happened. The proposed settlement between Google and Epic Games has been withdrawn, and Judge James Donato's remedies from the original antitrust verdict will now take effect. Google confirmed that beginning next week, it will begin distributing third-party app stores through Google Play itself. The remedy package, as Ars Technica reports, also includes lower service fees and a requirement that Google Play catalog apps be mirrored into rival stores. The case originated in 2020, when Epic added a direct-purchase path for V-Bucks in Fortnite in violation of Google's and Apple's payment rules.
Why it matters. Where Apple emerged from its Epic case with limited structural damage, Google's conduct — pressuring device makers away from pre-loading competing stores, and attempting to conceal that conduct — produced a materially more invasive remedy. Distributing rival stores inside Google Play is a first-of-its-kind structural mandate for a general-purpose mobile OS in the United States. It converts app-store distribution from a duopoly gated by OEM deals into an open shelf, at least on Android.
Who is affected. Every Android device shipped with Google Play — well over three billion active devices — will gain in-Play access to competing storefronts. Epic's own store, the Microsoft-backed mobile game store announced last year, Amazon Appstore, and Samsung Galaxy Store all become plausible mainstream distribution channels. Developers gain leverage in fee negotiations, with Google's take rate already forced lower under the remedy. For Google, Play Store services revenue — a high-margin contributor to the Services line — faces sustained pressure. Apple's App Store, insulated by its own case outcome, becomes the outlier as regulators globally note the asymmetric result.
What to watch next. Which stores Google actually surfaces first and how prominently they are displayed; whether developers begin routing higher-ARPU titles through non-Google billing to capture the fee delta; the response from EU regulators already enforcing Digital Markets Act obligations; and whether Apple faces renewed pressure to match the Android remedy.
Sources:
Today's three stories describe three different loci of leverage in the current tech stack. Thinking Machines is trying to relocate value from monolithic frontier models to customization infrastructure; TSMC's guidance shows that the physical layer beneath that competition is still tightening, not loosening; and the Google Play remedy is a reminder that the distribution layer, however entrenched, remains susceptible to structural intervention. Taken together, they suggest that the AI cycle's underlying economics — model access, silicon supply, and platform gatekeeping — are all in motion simultaneously, and rarely in the direction incumbents prefer.