Jonathan Frei

Buy it in ChatGPT: Instant Checkout and the Agentic Commerce Protocol We’re taking first steps toward agentic commerce in ChatGPT with new ways for people, AI agents, and businesses to shop together.

Is the future of E-commerce shopping without ever leaving your chatbot? The Agentic Commerce Protocol (ACP) by OpenAI seems like a step in that direction: openai.com/index/buy-it...

France will go to the polls once more. That won't resolve the underlying issues but maybe a different coalition can sell the necessary cuts to the French. A tried and tested method could be rabid populism that blames it all on the rich and foreigners t.co/IymNM2paHB

(PDF) Using Prisoner's Dilemma to Evaluate Corporate Tax Reform Proposals PDF | The game Prisoner’s Dilemma provides a helpful tool for understanding international tax competition. Nations would be best served by cooperating... | Find, read and cite all the research you nee...

or be faced by a shrinking tax base for government revenue if they do implement higher taxes on the asset-rich. regarding a third option: any global cooperation between all states to shift the tax burden will always fail due to the prisoners dilemma: www.researchgate.net/publication/...

Any extension of social benefits will either directly or indirectly (via inflation) be funded by the working class itself, making the wealth gap between the asset-rich and working-poor larger every year. either governments can successfully distract people from this fact

I hate to be a cynic - but governments can either choose between becoming populist or losing their mandate every other year. The reality is that national governments have largely lost their ability to tax rich individuals or corporations due to today's global mobility of capital.

Unstable governments are probably going to become a staple. especially in countries where the populace shuns economic reality and expects a welfare system that tax revenue cannot provide. proposed wealth taxes - when introduced unilaterally - will just lead to a lower tax base.

And here's the Sora 2 announcement - makes perfect sense now to create the content for the personalized feeds. And just putting an ad every now and then between the reels is a tried and tested monetization engine:
x.com/OpenAI/statu...

probably the first step towards ads integration once it's released for everybody. a personalized feed like insta / tiktok is the standard ads delivery mechanism. additional compute costs should be low because generated content can be shared with everyone at little extra cost. x.com/sama/status/...

And that is excluding all other costs like salaries, servers, bandwidth, data acquisition and web crawling, etc. Demand for LLMs is huge but current cost structure and pricing is unsustainable. I expect prices to increase and model size to decrease when funding gets tight.

So far we have also assumed that all GPUs run at 100% utilization - hardly realistic. More likely is 50%, hence halving our revenue to $71 and generating a net profit margin of ca. -400%. Therefore in order to break even with current models prices need to be at least 5x higher.

This means the training cost is about $0.125 additionally per 1M tokens: $15M training cost / 120T * 1M tokens. This increases our costs further: (369 + 142) * $0.125 by $64 from $288 to $352. With $141 in revenue we got a net profit of -$211.

These models also have a finite life-span; let's assume half a year until they are outdated. During that time the model will generate about 120T tokens - assuming openrouter.ai captures about 1% of global API usage and 50B Tokens are used by GPT-4o-mini in 1 week.

For a model like Llama we need to train on around 15 trillion tokens. 1 H100 can work on 1.9B tokens in about 20.5 days. Therefore to train the model in 20.5 days we need to rent 7680 H100 GPUs running in parallel. Total cost: 7680 * 20.5d * 24h * $4 = about $15M

This cluster can process ca. 369M input and ca. 142M output tokens (no parallelization) per hour. Meaning at GPT-4o-mini prices (comparable to Llama 3.1 70b) we get a revenue of $141 per hour. With costs at $288 we already lose $147 and haven't even included training costs.

For large deployments of hundreds of GPUs we have higher costs per hour than a single deployment (<$2) of one H100 GPU, due to additional equipment, networking etc. Roughly around 4$ per Hour and GPU, meaning it costs us $288 per hour (72 x $4) in our example.

For our reference calculation, let's use a pretty capable open-source model (Llama 3.1 70b) running on 72 H100 Nvidia GPUs and token input and output ratio of about 60% and 40% (input tokens usually keep growing during an ongoing conversation, hence ~60%).

Here are some back-of-the-envelope calculations why the current API prices for LLMs are only possible due to massive VC investments and corporate sponsoring. Long term prices either need to increase by 5-10x or model size needs to be reduced.

Interesting to follow LLM API usage via openrouter.ai Anthropic and Google seem to be clearly leading on the B2B front. While OpenAI captures the largest share of the B2C LLM market (according to app installations). Take a guess which market will have larger margins.

A stable CPI and a hotter than expected PPI means companies are so far not forwarding higher prices to consumers. Most likely to avoid drawing bad PR from the White House and being singled out. Unlikely to continue like this since it hurts shareholders. www.cnbc.com/2025/08/14/p...

XMR News: Inside Qubic’s Controversial 51% Takeover of the Monero Blockchain Qubic says it has achieved hashrate dominance over Monero, sparking concerns over the future of the network's decentralization.

www.coindesk.com/business/202...

How's the next Commissioner going to deal with the inevitable revisions? Will there be changes to the release schedule, quantity or quality of reports? Politicizing a supposedly independent bureau will turn investors away from the US - contrary to the administration's intentions.

Lower payout to miners due to halvings or stable supply lowering prices -> lower electricity consumption by network -> lower impact on regional prices -> geographical concentration in places with lowest energy prices -> only mining company w/ cheapest electricity survives

https://x.com/P3b7_/status/1955173413992984988

Don't know whether +51% control of the hash rate was achieved with malicious intent. But it's the most likely outcome for any PoW chain (unless prices keep rising forever). Monero has a fixed tail emission (0.6 XMR) - still couldn't prevent it. Here's why it's almost inevitable:

Fascinating to follow outrage over GPT-5 rollout. Model seems worse/less suitable for lots of people. I believe OpenAI tries to automatically reroute queries to bigger and smaller GPTs to optimize costs. Learnings: Users 1) are highly sensitive to model capability 2) want choice

in welchem Machtblöcke in gegenseitiger Konkurrenz stehen, Verhandlungen nicht mehr automatisch im Rahmen der WTO stattfinden und reine Interessenspolitik verfolgt wird. In solch einer Welt stehen nationale Interessen über Allem und man wird als kleines Land an die Wand gedrückt.

Deshalb ist die Wahl Trumps nicht der Ursprung dieser Entwicklungen - sondern eine logische Folge. Die Weltordnung von ca. 1990 - ca. 2020 konnte nie über längere Zeit Bestand haben, sondern war eine Ausnahme in der Weltgeschichte. Wir kehren nun zurück in den Normalzustand

haben Volkswirtschaften zwar enorm von der Öffnung der Märkte profitiert - diese Gewinne sind aber überdurchschnittlich bei den Eigentümern angefallen. Während es bei Arbeitern in vielen Fällen zu Job oder Wohlstandsverlusten kam.

Beides im Glauben, dass schlussendlich mehr Welthandel den Frieden und Wohlstand fördert. Beides ist nicht eingetreten - mit oder ohne Trump. So hat die hohe gegenseitige Abhängigkeit zwischen EU und Rus im Energiebereich keine Invasion in der Ukraine verhindert. Gleichzeitig

als einziger Hegemon übrig blieb und China noch keine Mittel hatte eine eigenständige Wirtschaftspolitik zu führen. In dieser Zeit etablierte der Westen unter US Führung verschiedene Institutionen die zum Ziel hatten den Welthandel mehr zu integrieren und zu ent-politisieren;

Es wird nicht ausgesprochen weil es ein Grundverständnis der souveränen und unabhängigen Schweiz in Frage stellt. Diese Unabhängigkeit war seit jeher ein Schein: Die Schweiz profitierte von einer einzigartigen Zeitperiode nach dem Zerfall der Soviet Union - in welcher die USA