People used to say that if you aren't paying for the product then you are the product. How do we think open weights will be different?
"You're the product" can only apply in cases where the model is being hosted by whoever developed it, so that they can capture the traffic.
Some of these open weight models have revenue sharing agreements between the hosting company (e.g. AWS) and the company that created it.
Apparently Ziphu make money from GLM in China by providing service and support for privately hosted instances, a bit like the RedHat business model.
[0]: https://www1.hkexnews.hk/listedco/listconews/sehk/2026/0831/...
The hard part about a wikipedia model, overall: it costs hundreds(?) of millions of dollars in compute to train these things. Will we be able to get enough donations to afford that?
https://meta.wikimedia.org/wiki/Wikimedia_Foundation_Annual_...
Annual Budget: $207.5M
https://news.ycombinator.com/item?id=49732270
I have been told pre-training is now much cheaper than post, but let's say $5M in total?
To be clear, I'd love a donation-funded open model that could compete with the big players. But I just don't see it on the horizon.
To me, a company like Nvidia may be the only rational company to make an open model. At least they can sell you compute. And they have. But AFAIK, they haven't come close to the frontier models and I haven't seen much movement from them in a while.
The position of the Chinese government is that AI should aid humanity and it shouldn't be "owned" by, in particular, a US tech company. That would be a strategic threat to China. So in classic Microsoft style (of commoditizing your complements), China promotes open weight models.
I believe that all these big AI companies are trying to cash out before the inevitable happens, which is their moat dries up. Their moat is the model and the hardware. Frontier models are typically viewed as having less than a year lead on China's best models. The hardware moat is access to NVidia's best chips. That one will last longer but is also at long-term risk.
Open weight models are like how Linux was when commercial *nixes were still common.
It sure feels like changing that is within the regulatory power of the state. The state can't compete because it isn't trying.
The larger problem of a collapse is the recession as the whole house of cards falls down, with a government unwilling or unable to stimulate the economy new-deal style.
It does not seem easy to figure out the ultimate beneficiary:
https://x.com/arpitrage/status/1227667060987023361?a=p-ErOXN...
This is the ANT bet:
1. With current technology, using more computer resources will result in a measurably smarter model.
2. ANT will have more compute resources than anyone else for the next 5 years.
3. The intelligence gap provides meaningful competitive advantage.
4. The regulatory, monetary, or military environments don’t stop ANT from doing its work.
If better would always mean "economically expedient", we all would fly on hypersound commercial jets.
If $2 model in the hands of good engineer can do the same as $20 model, anthropic will be soon killed by self-hosted open-weight models.
The question I’m asking myself is will a company with the smarter models out-compete companies paying for cheaper smart-enough models in the next few years.
Only a few companies will gain a competitive advantage paying for hypersonic travel compared to fast enough sub-sonic jet travel. I think it’s plausible that large portions of the economy will gain a competitive advantage by using the smartest models.
It is generally believed that the frontier models are 6-9 months ahead of the best open weight models. That's not really a moat. It's a creek.
Hardware is more of a moat for now but faces long-term issues such as:
1. Local hardware is rapidly improving. By 2030 I think you'll be able to run pretty hefty models on relatively cheap hardware and there's a business case for this too. For example, it's unclear if law firms using cloud models violates attorney-client privilege. The easiest solution is local models;
2. New NVidia hardware is going to immediately halve the value of billions in GPUs. We've got the R100/200 coming out now. What about the generation after that? What will it do to the B100/200s? I believe that hardware will ultimately come down to performance-per-Watt (or token-per-Watt if you prefer). I believe the depreciation time frame is going to be a lot shorter than some of these companies have factored in; and
3. By 2030 you'll probably have China making EUV chips commercially (IMHO) on homegrown tech. This will fundamentally change the AI hardware landscape as China rapidly catches up.
How much value can they extract from this delta.
> Major Investors Are Skeptical Anthropic IPO Can Achieve $1.5T Valuation (https://news.ycombinator.com/item?id=50386597)
In a way those equities act as stores of hope, where investors unconsciously agree to value them as if the ideal future path where all the promises are true is the one that will happen. Whereas they are more rational and have a balanced view towards other equities.
Not because it's sensible (it isn't), but because of how much hype is behind it, which is all that matters nowadays.
That still feels wildly overvalued. SpaceX went public at a $1.77 trillion valuation, and its P/E ratio is already a whopping 617.7. Anthropic doesn't even own data centers, rockets, a social media platform, or a profitable internet service.
On the flip side, their valuation isn't mostly based on one person who could leave or go (more) insane at any time.
As for not owning data centers, I think this might be Anthropic's saving grace actually. It seems like the price paid for renting is particularly cheap, possibly too cheap. And the meteor headed for all the hyperscalar investment is what happens when NVidia releases a new series of chips and the token-per-Watt numbers halve?
The big companies have very different views on the effective lifetime of GPUs, ranging from 4 years to 8 and possibly 10? years. I think 8+ is wildly optimistic. 10 year old AI chips already sell for scrap and I think that's going to get worse.
will the A100 be tomorrow's consumer favorite refurbished gaming GPU? will scrappy open weight non-profits be running them at their lowest power states to appear conscientious? who's lying to whom about the lifetimes of these things? do typical retail investors have a clue that these chips will be totally undesirable soon? do institutional ones? when can i afford a private evening locally vibe coding impossible-to-prove-i-vibe-coded-this solutions to code golf challenges from 1998‽ the questions abound
What assets do they have? A bunch of models that at best are 3 months before the competition? And based on their own reports can easily be copied with only a few days worth of queries?
And they barely own any hardware, everything is rented from competitors at a very high price.
Knowhow? Dont make me laugh. Google and deepseek are doing the innovation. When did Anthropic published a cool idea?
Infrastructure? They dont own any, and since their training and inference is notoriously inefficient they are bleeding money to data center providers.
Customers? Their customer list is shrinking by the day. Companies are reducing their token budget. Some completely moving to on-prem.
Mindshare? Your grandma probably knows about chatgpt but who is claude? Also everyone on the left and the right hate Anthropic right now.
Exactly what here is worth this much money??
I personally will pull out of any index fund that includes openai and anthropic. Let someone else loose their savings.