Open Source AI — September 25, 2026
Two stories today, both about the gap between AI's price tags and its real costs.
In Plain English: Two stories today, both about the gap between AI's price tags and its real costs. Oracle just told one of its data-center landlords it can't be held to its 2028 deadline after regulators blocked the gas pipeline the facility needed - a sign the physical buildout is running into real-world limits, not just financial ones. Meanwhile DeepSeek, one of the cheap open-weight labs that's supposed to be undercutting everyone, just tripled its prices and is still raising $7.5 billion - proof that even the "affordable" side of AI is getting more expensive to run. Today's Apple Angle explains why buying your own hardware sidesteps both problems.
Oracle Invokes Force Majeure on Its $165 Billion Project Jupiter Data Center After Regulators Reject Its Gas Pipeline
Oracle sent Blue Owl Capital a force majeure notice on the 2.5-gigawatt Project Jupiter campus in New Mexico, a contractual move that shields it from liability if the facility's planned 2028 online date slips - triggered after state regulators rejected the natural gas pipeline extension the site needed in July, following an earlier federal pipeline denial in March. Oracle says the project "remains on our planned schedule" and it is "fully committed to New Mexico," but is now soliciting 2GW of replacement renewable capacity to fill the gap. This is a direct follow-up to the debt-market stress on this same project noted in Tuesday's brief, where Project Jupiter's $18 billion in financing was already trading at 89-91 cents on the dollar - now the underlying construction timeline itself is in question, not just how the market is pricing it. For Lucien's thesis, this is prediction #1's cost-and-delay risk moving from a financing-market signal to an operational fact: permitting and power-grid constraints, not just capital costs, are the friction slowing the centralized buildout.
Infrastructure & Economics · Bloomberg · Sep 24, 2026 · Follow-up to Sep 22 coverage
DeepSeek's Revenue Hits a $1 Billion Run Rate - After It Tripled Its Own API Prices
DeepSeek's annualized revenue run rate jumped to roughly $1 billion, up from under $500 million just months earlier, after the company raised its API fees by 2.3 to 4.5 times depending on the model back in August - and founder Liang Wenfeng reportedly told investors that demand held up anyway. The Chinese lab, whose actual revenue for the first seven months of 2026 was about $71 million (versus roughly $7 million for all of 2025), is now finalizing a $7.5 billion funding round at a roughly $74 billion valuation. For Lucien's thesis, this complicates the simple version of prediction #2: DeepSeek is a genuine open-weight frontier lab, but its own hosted API just got several times more expensive, which is itself a data point for prediction #1's cost pressure - even leading "cheap" labs are raising retail prices as they chase revenue, which only strengthens the case for running the (still freely downloadable) weights yourself instead of paying anyone's API.
Infrastructure & Economics · The Next Web (via The Information) · Sep 24, 2026
The Apple Angle (Standing Perspective)
Both stories above show the same rented, metered model of AI running into friction - one from state regulators and grid limits, the other from a "cheap" open-weight lab quietly tripling its own prices. Apple's Mac Studio sidesteps both: configure one with its full 512GB of unified memory, roughly $9,500, and it holds an entire trillion-parameter open-weight model in memory on a single desktop machine - no pipeline permit needed, no API price hike to absorb, because there is no metered bill in the first place.
Matching that memory footprint on Nvidia's own workstation line takes five to six RTX Pro 6000 cards, somewhere in the $60,000-$75,000 range, while drawing roughly ten times the power. A new frontier-capable open-weight model - the kind DeepSeek, GLM and Kimi keep shipping every few weeks - keeps nudging the real competitive question away from "who trained the smartest model" and toward "who can afford to own the hardware that runs a nearly-as-good one, locally." Apple's MLX software and Thunderbolt 5's RDMA networking already let several Mac Studios pool memory into one larger effective machine, Nvidia has dropped NVLink from its own consumer workstation cards, and its DGX Spark borrows more from Apple's unified-memory design than the reverse. The honest caveat still holds: a Mac Studio serves one person running one model at a time, while the gigawatt-scale facilities behind today's Project Jupiter story exist to serve millions of concurrent users at once - Apple looks positioned to quietly own a valuable slice of AI compute, not to replace the data center outright.
Limited Edition Jonathan (Substack) · Perspective
Why This Matters
Today's two items pull the same thread from opposite ends - a hyperscaler's data center running into real-world permitting and power limits, and a leading open-weight lab discovering its own prices aren't as fixed as the "cheap alternative" story suggests. Both point the same direction the standing Apple Angle keeps naming: owning the hardware sidesteps regulatory delay and price hikes alike, which is exactly the kind of pressure prediction #3 expects to eventually push adoption toward local, on-prem inference.
The Daily: Open Source AI - a recurring research brief for Lucien Engelen
Book Lucien for your next event
Keynotes, masterclasses, panels and board-room sessions.