Digital Colliers Daily Briefing — May 24, 2026
The economics, interfaces, and infrastructure of AI all moved today. DeepSeek converted a promotional discount into a permanent price floor that undercuts every Western frontier lab; Google used its I/O stage to push Gemini deeper into media creation, agents, and productivity software; and SpaceX flew Starship V3 cleanly on its debut, lending engineering credibility to an IPO pitch that now hinges on putting data centers in orbit. Below, the details and what they imply.
1. DeepSeek locks in a 75% price cut on V4 Pro, redrawing the inference cost floor
What happened. DeepSeek will make permanent the steep discount on its flagship V4-Pro API that had been scheduled to expire May 31, according to Bloomberg via Techmeme. Input tokens will be billed at $0.435 per million and output at $0.87 per million — roughly a quarter of the prior list price. The company framed the move as a permanent reset rather than an extended promotion.
Why it matters. At these levels, V4-Pro sits an order of magnitude below the per-token rates of OpenAI's and Anthropic's flagship tiers and well below Google's Gemini Pro pricing for comparable context windows. For workloads measured in billions of tokens per month — retrieval pipelines, agent loops, code review, synthetic data generation — the math now favors DeepSeek by a margin large enough to override soft preferences for U.S.-hosted models in price-sensitive markets.
Who is affected. Developers and startups running high-volume inference, particularly outside the U.S., gain a materially cheaper frontier-class option. OpenAI, Anthropic, and Google face renewed pressure on API list prices that have so far held despite previous DeepSeek rounds. Inference-optimized chip and cloud vendors benefit from the elasticity this unlocks; on the other side, the labs whose unit economics depend on premium API margins must decide whether to match, segment, or differentiate on capability.
What to watch next. Whether U.S. labs respond with explicit price cuts (unlikely in the short term) or with quieter enterprise discounting and bundled-tier moves. Also worth watching: enterprise procurement language around Chinese-hosted models, and whether DeepSeek's permanent pricing holds once V4-Pro sees sustained production load.
Sources:
2. Google I/O: Gemini 3.5 Flash, an "anything-to-anything" Omni model, Spark agent, and Docs Live
What happened. Google used I/O to roll out a wide AI slate. Gemini 3.5 Flash launched as the first model in the new 3.5 family, positioned as a faster, cheaper tier ahead of a later Gemini 3.5 Pro release, according to Matt Wolfe's on-the-ground recap with the DeepMind team. Google also introduced Gemini Omni, a multimodal editing and generation model that DeepMind staff pitched as the "Nano Banana of video" — currently accepting video inputs, with audio and image inputs to follow. The Verge's hands-on calls the resulting deepfake-quality output "wild," noting how little effort realistic synthetic video now requires. Google additionally previewed Gemini Spark, a server-hosted autonomous agent meant to run continuously without local compute, and Google Docs Live, an AI voice-driven drafting tool that the Wall Street Journal reports will roll out this summer to AI Pro and Ultra subscribers.
Why it matters. Each piece targets a different competitive front. 3.5 Flash answers DeepSeek and other low-cost APIs on price-performance. Omni puts Google directly in the generative-video lane currently dominated by OpenAI's Sora and a handful of specialty startups, with distribution advantages no competitor can match. Spark is Google's response to the wave of open agent frameworks — what Wolfe characterizes as "Google's version of OpenClaw or Hermes" — but hosted, which sidesteps the local-runtime constraint that has kept consumer agent adoption thin. Docs Live brings voice-first drafting into the most widely used productivity surface on the web, paywalled behind the Pro and Ultra tiers Google has been pushing as its AI subscription business.
Who is affected. Workspace's hundreds of millions of users are the immediate distribution target for Docs Live. Video creators, advertisers, and the trust-and-safety apparatus around synthetic media all inherit Omni's downstream effects — The Verge's piece foregrounds how indistinguishable the output already is from genuine footage. Agent startups face a hosted competitor with Google-scale infrastructure behind it. And OpenAI, whose video and agent narratives have been central to its valuation story, now contends with a same-week counter across all three axes.
What to watch next. Pricing and rate limits for Gemini 3.5 Pro when it lands; provenance and watermarking commitments around Omni-generated video; Spark's task scope and reliability data once it's in broader hands; and whether Docs Live drives measurable Pro/Ultra conversion when it ships this summer.
Sources:
- The Only Important Announcements From Google I/O — YouTube · Matt Wolfe
- Hands-on with a pre-release build of Google Docs Live, an AI-powered voice tool for drafting documents, rolling out this summer to AI Pro and Ultra subscribers — Techmeme
- Google's new anything-to-anything AI model is wild — The Verge AI
3. Starship V3 flies clean on debut, anchoring SpaceX's orbital data center pitch
What happened. SpaceX launched the first test flight of Starship V3 from Starbase in South Texas at 5:30 pm CDT Friday, Ars Technica reports. The 408-foot, 33-engine stack cleared the tower, turned eastward over the Gulf of Mexico, and splashed down on target in the Indian Ocean roughly an hour later. The result is a notable contrast with the V1 and V2 debuts in 2023 and 2025, both of which broke apart on their inaugural flights.
Why it matters. A working V3 is the engineering foundation under the more speculative claims in this week's SpaceX IPO filing — including, per TechCrunch, the company's argument that space-based solar-powered data centers can generate "more than five-times the energy" of terrestrial ones and that "terawatt-scale annual AI compute growth" will outrun Earth-based power supply. Without a reliable heavy-lift platform, those projections read as marketing; with one, they begin to look like a roadmap, however aggressive.
Who is affected. SpaceX IPO investors now have a concrete data point to weight against the filing's longer-horizon claims. Hyperscalers running into permitting, grid, and water constraints — the NIMBY problem TechCrunch flags — gain a hypothetical alternative path, even if the economics remain unfavorable for the foreseeable future. Tesla's solar business, conspicuously absent from xAI's procurement and from the SpaceX filing's terrestrial energy story, sits awkwardly in Musk's reordered portfolio: xAI is buying natural gas turbines and Tesla Megapacks but not Tesla panels, while SpaceX positions solar as something to be done in orbit.
What to watch next. V3's second flight and the cadence of subsequent tests; any disclosure of payload mass to orbit and reuse turnaround targets; and whether SpaceX puts dollar figures or timelines on the orbital data center concept in IPO roadshow materials. The TechCrunch analysis is right to flag the unresolved questions — radiation-hardened silicon costs, distributed-training feasibility across satellites, launch energy versus terrestrial deployment — that V3's success does not, on its own, answer.
Sources:
- SpaceX's Starship V3—still a work in progress—mostly successful on first flight — Ars Technica
- Elon Musk has given up on solar power (on Earth) — TechCrunch AI
Today's three threads converge on the same underlying question: where the constraints on AI actually bind. DeepSeek argues the binding constraint is price and is willing to compress margins to prove it; Google argues it is product surface area and distribution, and is bundling models, agents, and voice into Workspace accordingly; SpaceX argues it is power and land, and is asking public markets to fund an orbital answer. Each bet implies a different shape for the industry by the end of the decade, and none of them are compatible with all of the others.