@EMostaque So this makes the cost of like for like intelligence cheaper (assuming Chips are effectively of equal level wrt inference etc) but doesn't necessarily affect how much intelligence will still be utilised at 'x' cost regardless because it has to be (Govts.... businesses)
When the Government Can Switch Off Your AI, Is It Really Yours …. but wait there’s more?
The Fable 5 and Mythos 5 shutdown raises a potentially existential threat to frontier models as we know it or were expecting to know it in the context of them being adopted widely at scale and ubiquitous in the future.
Anthropic pulled two of its most powerful models globally after a single US government directive. Not just restricted for foreign users, but Globally and with no warning, no SLA protection, no grace period for businesses mid-deployment the lights went out for paying customers worldwide for those best in class models. This was upshift to an already increasing scenario of nerfing models based on inputs which is essentially a form of user censorship. Even if you acknowledge and accept that in the context of improper prompts (such a bioweapon conversations), it opens up a whole pandoras box of end user trust and validity arguments to variable results from your AI solution firewalled by rules you don’t have any control over.
What strikes me most is how little attention is being paid to the second and third order consequences here.
If frontier models become subject to export controls at the level we've just seen, the case for making them the backbone of serious enterprise infrastructure starts to look very shaky. That this infrastructure can disappear overnight or even the broader concept of nerfing under the cloak of ‘we don’t think you should have access to this’ changes the fundamental usability, trust and value proposition frontier AI appeared to offer and would surely lead to a re-evaluation of investment and adoption.
The CNI question is also very interesting. If these (and future) models get some level of Critical National Infrastructure classification the accompanying oversight regime would likely be incompatible with the pace of development that makes frontier AI commercially valuable.
The innovation velocity that justifies the valuations gets constrained at exactly the point governments decide to protect it.
The immense CAPEX regime and forms of circular financial architecture holding this sector together deserves more scrutiny than it's getting. The Hyperscaler investments in Anthropic and OpenAI, the cloud-credit commitments, the CAPEX bets, all of it prices in unrestricted global deployment and large ROI. A world where models are subject to sudden export controls starts to look uncomfortable for that financial model.
Anthropic put it plainly themselves: if this jailbreak standard were applied consistently across the industry it would effectively halt all new frontier model deployments.
So will we be procuring the very best leading edge AI capability or procuring access to AI capability at the government's discretion? Those are very different (value) propositions.
#AI #Hyperscalers #Fable5 #CAPEX
@Jason US Govt action surely condemning frontier models to CNI and never realising all of that ROI > reduction in CAPEX > collapse of the circular funding cycle > a big deflation to what wasn't actually going to be a bubble!
@RaoulGMI@AndreasSteno Hyperscaler rotation ..... all that CAPEX funding something that the Govt can shutdown and rug you/your Company and RSI makes open source on your own hardware the only 'safe' way to go ..... leaving the Hyperscalers 💩
@JOBhakdi Yes it is Jo - rotation out of Hyperscalers indeed .... the AI bottlenecks trade for a while yet and one day Tesla will get above and stay $500... (don't mention the you know what!)
@WyckoffAnalysis@martypartymusic
divergence on timelines .....
Where does Clarity Act passing fall must surely be a fundamental pivot against the chart timeline to come?
Waymo are not only driving into flood zones, one way streets, down railroad tracks but now also into freeway construction zones.
The party trick is finding edge cases
If this was TESLA it would be global news...
List of WAYMO Incidents so far
1,429 (July 2021 – Nov 17, 2025 per NHTSA data analysis). Some sources put the cumulative figure around 1,790 as of March 2026. The number continues to grow as they log more miles.
Injuries/Fatalities (through late 2025):
2 DEAD
117 injuries
1 Dog Killed that we know of
@EMostaque All that aside though Emad - what/who/how can we trust? .... I can't get past that as an overriding factor for where I want to source my personal/families AI