AI CORNER
Tech stocks tumble as China’s DeepSeek sows doubts about AI spending Start-up’s model raises questions about need for huge western hardware investment
Thanks to David’s research, this blog has been documenting the stealth advances of Chinese AI models. And it’s not only DeepSeek.
David two weeks ago:
With the advent of chatgpt a few years ago, the research community was locked out of making progress because only hyperscalers could provide the compute necessary and the funds.
Chinese are demonstrating how it is possible to train models that rival the reasoning capabilities of the o1 series by using very small models at the very low cost of 450$ to train. Sky T1 was trained with synthetic data generated by Alibaba’s QwQ reasoning model which was then used to fine tune another Alibaba model: Qwen 2.5
Moreover, it will now be possible to do so on a 3000$ machine called “DIGITS” which was just announced by Nvidia (No cloud required).
He prompted Perplexity.ai on the significance of Chinese models:
Perplexity Pro costs $20/month. By comparison, ChatGPT Pro costs $200/month and Sam Altman claims they are losing money at that price. Understandable when your capital cost is in billions:
Perplexity allows users to chose among a number of reasoning models. Last week, David asked ChatGPT o1 model if Perplexity should replace it with DeepSeek r1. As David said, “gotta appreciate the honesty”:
All this to realize that the costs of AI models and its future applications is dropping fast. David says that “Alibaba is building opensource tools to develop these new agents at 50x less inference cost than western providers.”
Alibaba Cloud, the cloud computing arm of Alibaba Group Holding Ltd., has announced significant price cuts on its large-language models, a move that is set to revolutionize the AI landscape. The company has reduced the price of its Qwen-Long model by up to 97%, making it the most affordable large-language model in the market. Additionally, the Tongyi Qianwen series visual understanding models have seen an overall decrease of over 80%, with the Qwen-VL-Plus model’s price reduced by 81% and the Qwen-VL-Max model’s price slashed by 85%. These price cuts are expected to have a profound impact on user adoption, market share, and revenue growth for Alibaba Cloud’s AI services. (…)
The price cuts are also expected to stimulate market growth by encouraging more businesses to adopt AI technologies. This increased adoption can, in turn, drive innovation and the development of new AI applications, ultimately benefiting the entire market. The aggressive pricing strategy by Alibaba Cloud may prompt other cloud service providers to lower their prices as well, leading to a more competitive market landscape. This increased competition can result in better services and more affordable pricing for customers, fostering a healthier market environment.
Game changers!
Now it’s all mainstream:
Silicon Valley Is Raving About a Made-in-China AI Model DeepSeek is called ‘amazing and impressive’ despite working with less-advanced chips
(…) “Deepseek R1 is one of the most amazing and impressive breakthroughs I’ve ever seen,” said Marc Andreessen, the Silicon Valley venture capitalist who has been advising President Trump, in an X post on Friday. (…)
DeepSeek’s development was led by a Chinese hedge-fund manager, Liang Wenfeng, who has become the face of the country’s AI push. On Jan. 20, Liang met China’s premier and discussed how homegrown companies could narrow the gap with the U.S.
Specialists said DeepSeek’s technology still trails that of OpenAI and Google. But it is a close rival despite using fewer and less-advanced chips, and in some cases skipping steps that U.S. developers considered essential.
DeepSeek said training one of its latest models cost $5.6 million, compared with the $100 million to $1 billion range cited last year by Dario Amodei, chief executive of the AI developer Anthropic, as the cost of building a model.
Barrett Woodside, co-founder of the San Francisco AI hardware company Positron, said he and his colleagues have been abuzz about DeepSeek. “It’s very cool,” said Woodside, pointing to DeepSeek’s open-source models in which the software code behind the AI model is made available free. (…)
“The only strike against it is some half-baked PRC censorship,” said Woodside, referring to the People’s Republic of China, but he said this could be removed because other developers can freely modify the code.
DeepSeek said R1 and V3 both performed better than or close to leading Western models. As of Saturday, the two models were ranked in the top 10 on Chatbot Arena, a platform hosted by University of California, Berkeley, researchers that rates chatbot performance. A Google Gemini model was in the top spot, while DeepSeek bested Anthropic’s Claude and Grok from Elon Musk’s xAI. (…)
While DeepSeek’s flagship model is free, the company charges users who connect their own applications to DeepSeek’s model and computing infrastructure. An example is a business that wants to tap the technology to give AI answers to customers’ queries.
Early last year, DeepSeek cut its prices for this service to a fraction of what other vendors charged, prompting the industry in China to start a price battle.
Anthony Poo, co-founder of a Silicon Valley-based startup using generative AI to predict financial returns, said his company moved to DeepSeek from Anthropic’s Claude model in September. Tests showed DeepSeek performed similarly for around one-fourth of the cost.
“OpenAI’s model is the best in performance, but we also don’t want to pay for capacities we don’t need,” said Poo. (…)
DeepSeek said in a technical report that it used a cluster of more than 2,000 Nvidia chips to train its V3 model, compared with tens of thousands of chips for training models of similar size. A few U.S. AI specialists have recently questioned whether High-Flyer and DeepSeek are accessing computing power beyond what they have announced. (…)
DeepSeek said its model, designed to solve tricky math word problems and similar challenges, was comparable to OpenAI’s reasoning model o1 even though it omitted supervised fine-tuning and focused on reinforcement learning—essentially directed trial and error.
Jim Fan, a senior research scientist at Nvidia, hailed as a breakthrough the DeepSeek paper reporting the results. He said on X it reminded him of earlier pioneering AI programs that mastered board games such as chess “from scratch, without imitating human grandmasters first.”
Zack Kass, a former executive at OpenAI, said DeepSeek’s advances despite American restrictions “underscore a broader lesson: Resource constraints often fuel creativity.”
Bloomberg: (…) “DeepSeek shows that it is possible to develop powerful AI models that cost less,” said Vey-Sern Ling, managing director at Union Bancaire Privee. “It can potentially derail the investment case for the entire AI supply chain, which is driven by high spending from a small handful of hyperscalers.” (…)
Will DeepSeek Sink The Unsinkable Mag-7?
(…) On balance, we expect that the Mag-7 will deliver solid earnings, as suggested by the record high in their combined aggregate forward earnings of $500.1 billion during the week of January 24. The forward earnings of the S&P 493 was $1,819.3 billion that same week.
The new issue for the Magnificent-7 is whether DeepSeek will deep-six their AI aspirations.
This Chinese company recently shocked the tech industry when it reportedly spent only $5.6 million over two months to develop its latest LLM, which outperformed rival US LLMs from Meta and ChatGPT. The company kept costs down by using less powerful and cheaper Nvidia H800 GPU chips. Its LLM is available on an open-source basis.
This might be bad news for the Mag-7 that have plans to dominate the AI market with their (expensive) AI services. On the other hand, it might mean that AI systems will be more accessible and cheaper. If so, the best way to play AI might be the S&P 493 companies that will be cutting their costs and boosting their productivity using this new technology.
It might be good news for the Mag-7 that can learn from DeepSeek to design AI systems with cheaper GPUs. That would reduce their capital spending and boost their profits. It might not be a happy development for Nvidia.
U.S. Flash PMI: Output growth slows in January and price pressures rise, but employment jumps higher on sustained optimism
The headline S&P Global US PMI Composite Output Index fell from 55.4 in December to a nine-month low of 52.4 in January, according to the preliminary ‘flash’ reading, which is based on approximately 85% of usual survey responses.
The slowdown was centered on the services economy, where output rose at the slowest rate since last April, albeit sustaining the sector’s expansion into a twenty-fourth successive month. Manufacturing output meanwhile rose marginally, representing an improvement on the declines recorded over the previous five months.
Changing activity levels reflected varying demand conditions. While inflows of new business into the service sector remained robust, the rate of increase waned to a three-month low amid the first fall in overseas (export) orders since last June. Especially adverse weather was reported as a dampener of activity in some companies.
Manufacturers meanwhile reported the first, albeit very modest, rise in new orders for seven months, reflecting improved domestic demand and a softening rate of loss of export orders.
Optimism about the coming year continued to run at an elevated level. Measured across goods and services, firms’ expectations of their output in the next 12 months was unchanged in December, thereby remaining the joint-highest since May 2022.
Service sector confidence lost some of the shine from December’s one-and-a-half year high, but remained the second-highest recorded over the past year. Manufacturing confidence meanwhile surged higher, reaching the highest since March 2022 after posting the largest monthly improvement since November 2020.Uncertainty in the lead up to the Presidential Election has been replaced with optimism about the future, notably among manufacturers, according to anecdotal evidence provided by survey respondents. Looser regulation, lower taxes and heightened protectionism were all widely cited, alongside a broader sense of improving economic conditions in the year ahead under the new administration.
However, some companies express concern over the potential for policies such as tariffs to disrupt supply chains and impact sales, or stoke inflation. Others cite concerns over the strong dollar, high prices and the possibility of policymakers taking a more hawkish stance toward interest rates than previously anticipated.
Optimism about the year ahead was matched by a jump in hiring. Employment rose in January at the fastest rate for two-and-a-half years, up for a second successive month after four months of job shedding. The improvement was led by a surge in service sector hiring, where jobs were added at the sharpest rate for 30 months, though manufacturing payroll growth also edged up to a six-month high. The latter remained modest, however, reflecting ongoing cost concerns at producers amid low sales. Firms more broadly also continued to report ongoing issues with poor staff availability.
Inflationary pressures meanwhile intensified in January. Both input costs and average selling prices rose at the fastest rates for four months, the rate of inflation of the latter now having increased for two successive months.
Factory input prices rose at the steepest rate since last August, generally linked to supplier-driven raw material price increases. Growth of service providers’ costs also revived after having cooled to a ten-month low in December, rising at the fastest rate for three months amid increased staff costs and rising material prices.
Higher costs were passed on to customers, with average prices charged for services rising at the fastest rate since last September. An even larger rise was reported for goods, the rate of inflation of which hit a ten-month high.
The S&P Global Flash US Manufacturing PMI rose from 49.4 in December to 50.1 in January, signaling a marginal improvement in business conditions within the goods-producing sector and a contrast to the deteriorations seen over the prior sixth months.
Factory production rose marginally, increasing for the first time in six months, with new orders also returning to modest growth after six months of decline. Employment increased for a third successive month, the rate of job creation the highest since July.
Suppliers’ delivery times meanwhile lengthened for a fourth straight month, adding support to the PMI (longer lead-times often indicate busier supply chains), albeit slightly less so than in December. However, inventories fell at the steepest rate for 17 months, acting as a drag on the PMI, though this in part reflected higher than anticipated use of inputs in production rather than cost-focused destocking.
Commenting on the flash PMI data, Chris Williamson, Chief Business Economist at S&P Global Market Intelligence said:
“US businesses are starting 2025 in an upbeat mood on hopes that the new administration will help drive stronger economic growth. Rising optimism is most notable in the manufacturing sector, where expectations of growth over the coming year have surged higher as factories await support from the new policies of the Trump administration, though service providers are also entering 2025 in good spirits.
“Although output growth slowed slightly in January, sustained confidence suggests that this slowdown might be short-lived. Especially encouraging is the upturn in hiring that has been fueled by the improved business outlook, with jobs being created at a rate not seen for two-and-a-half years.
“However, rising price pressures are a concern, with companies reporting supplier-driven price hikes as well as wage growth amid poor staff availability. Higher input cost and selling price inflation was broad-based across goods and services and, if sustained, could add to worries that a combination of robust economic growth, a strong job market, and higher inflation could encourage a more hawkish policy approach from the Fed.”
“Adverse weather” slowed booming services demand but service companies hired at “at the sharpest rate for 30 months” amid “robust new orders”. Meanwhile, manufacturers reported the “first, albeit very modest, rise in new orders for seven months” and payroll growth edging up to a six-month high.
Good news for the Fed on the labor market.
But the bad news is “intensified inflationary pressures” as prices charged for services rose “at the fastest rate since last September” and “an even larger rise was reported for goods, the rate of inflation of which hit a ten-month high.”
Not a hint of productivity effect there…
Recent Fed regional surveys were not as upbeat:
- The January NY Fed surveys of manufacturing and non-manufacturing activity remained fairly subdued on all fronts.
- The January Philly Fed manufacturing survey was very strong on all aspects while its non-manufacturing was soft.
- Kansas City manufacturing was rather soft.
- The December Richmond Fed manufacturing survey was also soft but services were upbeat though tempered.
- The December Dallas Fed factory indices were also generally neutral while service sector activity was good but not inflationary.
Much has been said and written about the rise in U.S. productivity (unseen in most other economies) but, in reality, the sharp slowdown in employment costs in the last 2 years has only brought the ECI back to its 2012-2019 trend as Ed Yardeni illustrates:
The inflation genie is not completely back into the bottle, is it?
The next JOLTS report will likely decline back to trend but Indeed job postings have perked up recently (through Jan. 17).
New job postings might also have changed trend lately:
The U.S. consumer is in great shape and feels giddy:
Very giddy:
January U.S. Light-Vehicle Sales Continue Q4-2024’s Growth
While demand in the middle of the month was negatively impacted by extreme weather conditions across most of the country, with a week remaining in January there is upside to the outlook. On the flipside, there could be pause among some consumers, as they wait to see how the apparent revamping of federal policies and institutions by the new administration plays out. Regardless, sales are tracking to their fourth straight increase in January.
Almost back to pre-pandemic levels:
Imagine if vehicle production recovers.
Existing-Home Sales Ascended 2.2% in December
(…) Total existing-home sales – completed transactions that include single-family homes, townhomes, condominiums and co-ops – elevated 2.2% from November to a seasonally adjusted annual rate of 4.24 million in December. Year-over-year, sales swelled 9.3% (up from 3.88 million in December 2023).
This was the third consecutive year-over-year increase after declining YoY every month for over 3 years. Total housing inventory registered at the end of December was 1.15 million units, down 13.5% from November but up 16.2% from one year ago (990,000).
Imagine if the housing market recovers.
Slowly getting used to the old normal 6-7% mortgages.
The average 30-year fixed mortgage rate has been above 6% since September 2022 and above 7% on and off since October 2022. The daily measure by Mortgage News Daily is today at 7.11%. Freddie Mac’s weekly measure, released yesterday, of the average 30-year fixed mortgage rate was 6.96%.
The real estate industry has now given up waiting for mortgage rates to plunge to wherever and is encouraging sellers and buyers to get used to “a new normal of mortgage rates between 6% and 7%,” as the NAR had put it, which are the old normal rates that prevailed before the money-printing era started in 2009..
The CEO of Fannie Mae, the largest Government Sponsored Enterprise that buys and guarantees mortgages, also encouraged buyers, sellers, and everyone in the industry to get used to these 6% to 7% mortgage rates.
Before the money printing era, the average mortgage rates had been well above 5%. The Fed’s QE and zero-interest-rate policy, which started in 2008 and, with some interruptions, finally ended in 2022, had created an anomaly:
If housing recovers, where will the workers come from?
Data: Associated Builders and Contractors. Chart: Axios Visuals
The construction industry needs to attract 439,000 new workers this year to meet [current] demand, otherwise costs will rise — putting some projects out of reach — per projections from the Associated Builders and Contractors trade group out this morning.
Increased immigration during the Biden administration was a boon for the construction industry, which is perennially short on workers, but with the Trump administration cracking down on migration, progress could reverse.
- The issue takes on new urgency as swaths of Los Angeles need to be rebuilt in the aftermath of devastating fires.
- A coming surge in data center construction nationwide will also require resources.
Immigrants make up about 26% of the construction workforce, per census data cited by Pew Research Center last fall.
- The construction industry also employs the largest share of undocumented immigrant workers, among all other industries.
- An estimated 13% of construction workers are undocumented, per Pew. (Axios)
The total U.S. labor force grew 8.5 million since late 2020, 6 million of which were new immigrants, not only contributing to labor supply but also presumably to keep average wages lower than otherwise, not only in construction.
China’s Economy Stumbles in Sign Rebound Hinges on More Stimulus Manufacturing PMI unexpectedly fell to lowest since August
Factory activity shrank in January after three months of expansion, with the manufacturing purchasing managers’ index falling to 49.1, the lowest since August. The non-manufacturing gauge for construction and services dropped to 50.2, just above the 50-mark that separates growth and contraction. (…)
Both production and new orders fell to a five-month low, according to the PMI data. In a sign of weak global demand, new export orders dropped to the lowest since February.
EARNINGS WATCH
78 companies in the S&P 500 Index have reported earnings for Q4 2024. Of these companies, 78.2% reported earnings above analyst expectations and 14.1% reported earnings below analyst expectations. In a typical quarter (since 1994), 67% of companies beat estimates and 20% miss estimates. Over the past four quarters, 78% of companies beat the estimates and 17% missed estimates.
In aggregate, companies are reporting earnings that are 8.6% above estimates, which compares to a long-term (since 1994) average surprise factor of 4.2% and the average surprise factor over the prior four quarters of 6.6%.
Of these companies, 61.5% reported revenue above analyst expectations and 38.5% reported revenue below analyst expectations. In a typical quarter (since 2002), 62% of companies beat estimates and 38% miss estimates. Over the past four quarters, 62% of companies beat the estimates and 38% missed estimates.
In aggregate, companies are reporting revenues that are 1.0% above estimates, which compares to a long-term (since 2002) average surprise factor of 1.3% and the average surprise factor over the prior four quarters of 1.2%.
The estimated earnings growth rate for the S&P 500 for 24Q4 is 10.4%. If the energy sector is excluded, the growth rate improves to 13.8%.
The estimated earnings growth rate for the S&P 500 for 25Q1 is 11.5%. If the energy sector is excluded, the growth rate improves to 13.1%.
The estimated revenue growth rate for the S&P 500 for 24Q4 is 4.2%. If the energy sector is excluded, the growth rate improves to 4.9%.
The 78 companies that have reported had earnings growth of 25.8% on revenues up 4.5%. The surprise factor is +8.6%.
After Q3, the first 71 companies reporting had earnings growth of 7.8% on revenues up 4.6%. The surprise factor was +6.4%.
FYI:
While
Trump Tariffs Are a Wealth Killer Two centuries of experience proves the economic foolishness of taxing imports.
From Andy Kessler in the WSJ:
(…) In a populist bid to protect our dwindling manufacturing workforce, economically clueless Trump whisperers push tariffs. A select few workers may be helped, but most Americans will be worse off—though I doubt we’ll see riots protesting higher prices on made-in-China Gap clothes, Barbie dolls or Hush Puppies shoes. But new Trump tariffs will raise prices and restrict other countries from affording our high-margin exports—drugs, phones, planes and many software and artificial-intelligence services. That’s dumb. (…)
Notice how everything is now either a national-security concern or an emergency. Tariffs on Canadian bacon for national defense? A national emergency? (…)
Apologists try to rationalize tariffs as a negotiating tactic, violating the first rule of negotiating by even mentioning it. Tariffs are coming. Expect retaliation and inflation—precisely why tariffs don’t work. (…)
A 2021 study by Oxford Economics and the U.S.-China Business Council showed the first-term Mr. Trump’s tariffs and trade policies destroyed 245,000 jobs. The Tax Foundation estimates Trump-Biden tariffs reduced long-run gross domestic product by 0.2%—roughly $58 billion annually. On the flip side, the Peterson Institute for International Economics estimates that free trade since 1950 has cumulatively boosted the U.S. economy by $2.6 trillion, or $19,500 a household. Why go backward? Congress should reclaim its tariff power.
Instead, the backroom begging will start for tariff exemptions—machinery, certain pharmaceuticals, school pencils, cobalt for electric-car batteries, Nike Kobe 5 Protro “Year of the Mamba” sneakers—a lobbyist’s paradise. Free trade, not politicians, is best at allocating resources. Protectionism and mercantilism in the form of tariffs and subsidies, like the British Corn Laws, are inefficient, unproductive, corruption-inducing and wealth-destroying. That won’t make America great again.
FYI #2:
Source: Visual Capitalist