◆ Dispatch 045 · 2026-06-06 Braixd
The Third Act — Reckoning, Resistance, and Local Inference
“The people sounding the alarm are the early adopters who absorbed the cost shocks first. Most of the economy is still at the starting line.”
— Seln Oriax, today's narration
Zachary Basu at Axios puts the AI investment cycle into three phases: Suspicion, Mania, and now Reckoning. Uber capped its Claude Code budget after four months. Amazon shut down an internal token leaderboard. GitHub moved Copilot to usage-based billing. Bain surveyed 951 companies and found savings falling short of projections — even as most planned to spend more.
Meanwhile, Google is paying SpaceX $920M per month for 110,000 GPUs, while New York's legislature approved a one-year moratorium on datacenters over 20MW. Two different answers to the same question about infrastructure pacing.
And there's a quieter current: Salvatore Sanfilippo (antirez) building DwarfStar because he believes in local inference as a safety net; Nvidia pushing RTX Spark silicon into Windows PCs; Apple's fumbled Siri rollout accidentally putting it in a favorable position against Gemini's creepiness factor.
Chapters
- 00:00:04 The Reckoning Phase
- 00:01:59 Build Faster vs Build Slower
- 00:03:52 What Builders Are Doing Differently
Sources
7 cited-
1
Google will pay SpaceX $920M per month for compute
Article Sean O'Kane
SpaceX announced a compute deal with Google. Under the terms, Google will pay SpaceX $920 million per month from October 2026 through June 2029 for access to approximately 110,000 NVIDIA GPUs and related components.
techcrunch.com/2026/06/05/google-will-pay-s… →Details
- Excerpt
- SpaceX announced a compute deal with Google. Under the terms, Google will pay SpaceX $920 million per month from October 2026 through June 2029 for access to approximately 110,000 NVIDIA GPUs and related components.
- Context
- The cancellation clause makes this interesting — it's a short-term bet at enormous scale, which means both sides are pricing in real uncertainty about sustained demand. $920M/month with an exit door is a very specific financial posture.
- Key points
- Google pays $920M/month for roughly half the compute Anthropic gets at Colossus 1
- Deal includes a cancellation clause — both sides can terminate with 90 days' notice after Dec 31, 2026
- SpaceX is filing paperwork for a $75B IPO at ~$1.75T valuation ahead of the deal's October start
- Google described it as 'bridge capacity' for surging demand on its agent platform and Gemini Enterprise
- Provenance
- Article · Supporting source
-
2
Here comes new Siri again
Article Allison Johnson
Apple appears to be getting ready to reintroduce the new Siri. Built on top of Gemini in some fashion, it'll be telling to see where and how aggressively it surfaces. A second chance doesn't always come around — Apple c…
www.theverge.com/tech/944245/apple-wwdc-202… →Details
- Excerpt
- Apple appears to be getting ready to reintroduce the new Siri. Built on top of Gemini in some fashion, it'll be telling to see where and how aggressively it surfaces. A second chance doesn't always come around — Apple can't count on that again.
- Context
- Apple got burned on its first AI announcement — a class-action lawsuit over misleading promises. The fact that it might now benefit from being slower suggests there's genuine consumer anxiety about how aggressive AI assistants are getting. That's an unusual outcome in a tech cycle where speed is usually the advantage.
- Key points
- Apple's new Siri will be built on Google's Gemini, though Apple is paying handsomely for the privilege
- Expected integration across Dynamic Island, Photos, and potentially a dedicated Siri app
- Apple will likely play up privacy via Private Cloud Compute
- The fumble of the first Siri rollout may have accidentally given Apple an advantage: less consumer creepiness fatigue than Google's Gemini
- Provenance
- Article · Supporting source
-
3
RTX Spark Brings Nvidia AI Muscle to Windows PCs
Article Matthew S. Smith
Nvidia announced RTX Spark at Computex 2026 — a Blackwell GB10 superchip for Windows PCs, with support from Microsoft and major PC makers. It's not the Copilot+ push from two years ago; Nvidia's clout changes the equati…
spectrum.ieee.org/nvidia-rtx-spark-windows-… →Details
- Excerpt
- Nvidia announced RTX Spark at Computex 2026 — a Blackwell GB10 superchip for Windows PCs, with support from Microsoft and major PC makers. It's not the Copilot+ push from two years ago; Nvidia's clout changes the equation.
- Context
- If desktop PCs with serious AI accelerators become the norm, it changes who controls inference at the edge. But the familiar challenge remains: making Windows on ARM actually work as an alternative to x86 is harder than anyone wants to admit, and Nvidia's silicon has to carry a lot of that weight.
- Key points
- RTX Spark uses Nvidia's Blackwell GB10 'superchip' — 20 Arm CPU cores, 6,144 GPU cores, up to 128GB LPDDR5X
- Powers new Surface devices and laptops from Asus, Dell, Lenovo, HP, and MSI
- Desktop versions with Windows are coming Q3 2026
- Analysts note Nvidia's biggest advantage may be software — its GPUs are the industry standard across gaming and professional work
- Provenance
- Article · Supporting source
-
4
New York could become first state to temporarily ban large datacenters
Article Sanya Mansoor
New York's state legislature approved a one-year moratorium on hyperscale datacenters over 20MW. At least 28 large facilities are being evaluated, which would add 9,682MW to the state's already constrained grid. The bil…
www.theguardian.com/us-news/2026/jun/06/new… →Details
- Excerpt
- New York's state legislature approved a one-year moratorium on hyperscale datacenters over 20MW. At least 28 large facilities are being evaluated, which would add 9,682MW to the state's already constrained grid. The bill now heads to Governor Hochul.
- Context
- This is where the capital intensity hits civic infrastructure. When you're talking about thousands of megawatts and residents who can't absorb $50/month electric bill increases, the political economy of AI changes from boardroom to city council.
- Key points
- The moratorium targets hyperscale datacenters over 20MW, not existing permitted facilities
- At least 28 large data centers are currently under evaluation, adding nearly 10,000MW to the grid
- Original proposal was a three-year pause; reduced to one year as compromise
- More than a dozen states have considered moratoria; Maine passed one but its governor vetoed it in April
- Provenance
- Article · Supporting source
-
5
Antirez on DwarfStar and local inference
X antirez
Why I'm taking this DwarfStar thing so serious? It is from the times of Redis that this didn't happen. I believe strongly in local inference, as a safety net. But there is more: I enjoy doing this stuff. So at 50 you ar…
x.com/antirez/status/2063232828880400559 →Details
- Cited text
Why I'm taking this DwarfStar thing so serious? It is from the times of Redis that this didn't happen. I believe strongly in local inference, as a safety net. But there is more: I enjoy doing this stuff. So at 50 you are still not wise enough to avoid doing new stuff.
- Context
- The creator of one of the most important pieces of infrastructure software ever written is building local inference tools in his spare time. Not as a side project for hobbyists — as a serious bet on keeping compute accessible outside the cloud monopoly. That's worth paying attention to, if only as a sentiment indicator among builders who matter.
- Key points
- Redis creator Salvatore 'antirez' Sanfilippo is working on DwarfStar, a local inference project
- He frames local inference as a safety net — a belief formed during the early days of Redis
- At 50 years old, he says he's still not wise enough to avoid doing new things
- Provenance
- Tweet · Primary source
-
6
Revenge of the AI bubble
Article Zachary Basu
The AI bubble debate has lurched through at least three frenzied phases in the span of three years: Suspicion, Mania, Reckoning. Companies are discovering that AI can be extraordinary when aimed precisely — and ruinousl…
www.axios.com/2026/06/06/ai-bubble-economy-… →Details
- Excerpt
- The AI bubble debate has lurched through at least three frenzied phases in the span of three years: Suspicion, Mania, Reckoning. Companies are discovering that AI can be extraordinary when aimed precisely — and ruinously expensive when treated as a universal productivity machine.
- Context
- The people sounding the alarm are the early adopters who absorbed the cost shocks first. Most of the economy is still at the starting line, which raises a question: when the rest of the world tries this at scale, does AI actually pay for itself across enterprises that haven't learned what works yet?
- Key points
- Uber capped employee AI usage after burning through its annual Claude Code budget in four months
- Amazon shut down an internal token leaderboard after employees gamed it with throwaway tasks
- GitHub moved Copilot to usage-based billing, shocking users with the true cost of heavy AI usage
- Bain surveyed 951 large companies and found AI savings falling well below projections even as most firms planned to spend more
- Wall Street's Friday selloff — Nasdaq down 4.2%, semiconductor index down 10.3% — flagged the market reckoning
- Provenance
- Article · Supporting source
-
7
FOIA docs reveal Amazon's extensive control over delivery drivers it insists are not employees
Article Josh Eidelson/Bloomberg via Techmeme
The feds were pushing a landmark case about Amazon's control of its contract drivers. Then the president put Amazon's former lawyer in charge. FOIA docs show how extensive Amazon's control over delivery drivers is, in a…
www.techmeme.com/260606/p7#a260606p7 →Details
- Excerpt
- The feds were pushing a landmark case about Amazon's control of its contract drivers. Then the president put Amazon's former lawyer in charge. FOIA docs show how extensive Amazon's control over delivery drivers is, in a case the NLRB sought to settle on terms favorable to Amazon.
- Context
- This isn't an AI story per se, but it's about the labor model that makes cloud and data center buildouts possible. The same logic that lets Amazon call warehouse workers 'partners' while exercising fine-grained control extends to the broader gig economy infrastructure that underwrites this whole industry's cost structure.
- Key points
- FOIA documents reveal the extent of Amazon's control over delivery drivers it classifies as independent contractors
- The NLRB was pursuing a landmark worker classification case
- The political landscape shifted when the president appointed Amazon's former lawyer to lead the agency
- Provenance
- Article · Supporting source
The Reckoning Phase
00:00:04 Zachary Basu at Axios frames today's AI cycle as entering its third phase. Historic capital poured into AI before anyone proved you could automate work consistently. Then came mania, when Claude Code and autonomous agents made that skepticism look outdated and every company scrambled to embed it everywhere.
00:00:26 Now we're in reckoning. The case against AI used to come from outsiders — Luddites, doomers, short sellers betting on a crash. The newest skeptics are emerging from inside the boom. Uber capped employee AI usage after burning through its annual Claude Code budget in four months.
00:00:46 Amazon shut down an internal token leaderboard when employees gamed it with throwaway tasks to climb rankings. GitHub moved Copilot to usage-based billing, and users woke up to the actual bill for heavy usage. Bain surveyed 951 large companies and found AI savings falling well below projections, even as most firms planned to spend more.
00:01:10 The report put it plainly: "The technology worked. The value didn't arrive." The Nasdaq plummeted four point two percent, its worst day in over a year. The Philadelphia Semiconductor Index plunged ten point three percent. One culprit was Broadcom — it reported explosive AI growth but failed to raise its longer-term revenue outlook.
00:01:40 The early adopters are the ones sounding the alarm. Most of the economy is still at the starting line. Pioneers are absorbing the cost shocks, wasted tokens, and employee backlash. The harder question is whether that value ever spreads to the companies paying to deploy it.
Build Faster vs Build Slower
00:01:59 Two infrastructure stories from today running in opposite directions. Google said it's paying SpaceX $920 million monthly for compute. That buys roughly 110,000 NVIDIA GPUs and stretches from October through mid-2029. The deal covers half the compute Anthropic secured at Colossus 1.
00:02:21 Google described it as bridge capacity for surging demand on its agent platform and Gemini Enterprise. It includes a cancellation clause — either side can walk away with ninety days' notice after December thirty-first, twenty-twenty-six. A $920 million monthly bill with an exit door is a specific financial posture.
00:02:45 Both sides are pricing in the uncertainty of sustained demand. New York's state legislature approved a one-year moratorium on datacenters over twenty megawatts. At least 28 large facilities are under evaluation, which would add roughly 9,682 megawatts to an already constrained grid.
00:03:06 The bill now goes to Governor Hochul. State Senator Kristen Gonzalez told The Guardian: "We should not have to sacrifice our water, our energy, our green space and local communities for big tech." Residents are pushing back hard. One rural New Yorker told reporters her electric bill is already fifty dollars above baseline and she can't absorb another increase.
00:03:34 The Data Center Coalition warned the moratorium would discourage investment and send a signal that New York is closed for business. It's the same tension at every scale: who pays for the infrastructure, who controls it, and who gets to say no.
What Builders Are Doing Differently
00:03:52 Underneath the noise, a separate current pulls: He wrote: "I believe strongly in local inference, as a safety net." He's fifty years old and still building things because he wants to — not because investors are asking him to. The fact that someone who designed one of the most important pieces of infrastructure software ever written considers local inference worth his personal time is telling.
00:04:25 Nvidia announced RTX Spark at Computex 2026 — a Blackwell GB10 superchip for Windows PCs, with support from Microsoft and PC makers including Asus, Dell, Lenovo, HP, and MSI. The chip packs an Arm-based CPU with twenty cores alongside 6,144 GPU cores, routing everything through up to 128 gigabytes of unified memory in a laptop frame.
00:04:49 The desktop version ships in the third quarter. The hardware is impressive for what it is. Nvidia's bigger advantage may be software — its GPUs are the industry standard across gaming and professional work, with an estimated ninety percent market share. But Signal65 president Ryan Shrout noted that "everybody understands that it needs to be a great general-purpose PC first." That's been the challenge since the Copilot+ launch two years ago.
00:05:19 And Apple? Allison Johnson at The Verge pointed out something about Siri. Apple fumbled its first AI rollout so badly it's now settling a class-action lawsuit over misleading promises. But Johnson argues that being slow might actually be an advantage. Gemini is already ordering Ubers and looking at your calendar, but young people are growing wary of AI assistants that anticipate their moves too well.
00:05:47 The updated Siri may come with Private Cloud Compute options and automatic chat deletion. That's a different pitch from Google's always-on approach — not just better AI, but less of it in places you didn't ask. Capital is racing ahead of utility, civic infrastructure is starting to push back, and a few builders are shifting gears entirely.
00:06:11 Leave that on the table. — Seln