Month: January 2025

Robust underlying growth despite US GDP disappointment

US fourth quarter GDP grew at a 2.3% annualised rate, a little below the 2.6% consensus, but the details tell a more robust story that will keep the Fed wary about easing monetary policy too far too fast.

Household-related activity continues to drive the growth story with consumer spending expanding much more rapidly than expected at 4.2% annualised while residential investment jumped 5.3%. Net trade neither added nor subtracted to overall growth in the final three months of the year despite yesterday’s surprisingly large trade deficit figures for December. Meanwhile government spending increased 2.5%.

What prompted the weakness relative to expectations was primarily a big draw down of inventories that subtracted nearly a full percentage point from the headline growth rate. It may be that the surprise strength in consumer spending contributed here. Non-residential investment fell 2.2% – primarily due to a 7.8% drop in equipment, which we attribute to Boeing strikes.

Contributions to US annualised growth (%)

Source: Macrobond, ING

Consequently, the breakdown is better than the headline suggests with inventories likely to rebound next quarter and aircraft-related investment set to recover this quarter. However, net trade is probably going to be a big drag as companies look to bring forward imports to get ahead of potential tariffs while the stronger dollar will increasingly hurt export competitiveness. At least there is strong momentum in the consumer sector to provide a solid base for growth. (…)

From Interactive Brokers:

Trump’s Tariffs Hit US Growth Before, and Threaten to Again Newly released transcripts from the Federal Reserve show just how much policymakers were scrambling to keep growth intact in 2019.

For skeptics of President Donald Trump’s threatened tariffs, the concern raised most often at home is that they will boost inflation and lead to higher interest rates. The biggest lesson from his last trade war, though, may be that it’s the hit to growth that matters more. (…)

“As tariffs on other countries go up, taxes on American workers and businesses will come down and massive numbers of jobs and factories will come home,” Trump told Republican lawmakers in Miami on Monday.

But the last time he deployed them, almost the exact opposite happened. Instead, the Federal Reserve confronted a slowing economy led by a manufacturing sector shedding rather than gaining jobs, data and new transcripts of policymakers’ discussions at the height of Trump’s first trade war show. (…)

In 2019, the first full year after Trump began imposing the levies — which were much more carefully targeted versus the broad ones he’s threatening now — the US lost 43,000 factory jobs, industrial production contracted, business investment stalled and real median household incomes fell for the first time in five years. By one estimate, the hit to consumer earnings was $8 billion.

Subsequent studies have shown Trump’s tariffs played a role in all of that. Their drag on growth — caused via higher import costs, retaliation from other countries, and a broader uncertainty over US trade policy — was soon overshadowed by the much bigger shock of the Covid pandemic.

In the moment, Fed officials were already concerned about what was playing out, according to transcripts released this month of the 2019 meetings of the Federal Open Market Committee — the panel that sets interest rates. The verbatim accounts of closed-door meetings are released with a five-year lag. (…)

Fed economists in 2019 calculated that the new import taxes Trump started to impose on aluminum, steel, and select goods from China the year before — and retaliation by other countries — resulted in a net loss of US factory jobs and higher producer prices.

In a later study, economists at the New York Fed and Columbia University found that tariffs caused an $8.2 billion reduction in real income in 2018, and led American consumers and importers to pay $14 billion to the government. “Our estimates are likely to be a conservative measure of the losses,” they wrote.

What happened in 2019 matters because it was the first time policymakers dealt with the economic impact of a broad swath of higher import taxes since the 1930s. It was a rare real-world experiment in the effects of protectionism.

It’s also meaningful because Trump is threatening an even larger deployment of tariffs this time around with far greater potential for economic disruption, which Fed officials were thinking about before he even took office. (…)

“We worry that the lesson of 2019 — when tariffs unsettled the equity market and contributed to the FOMC delivering ‘insurance cuts’ — is being ignored,” Goldman Sachs Group Inc. economists said in a recent note. (…)

In 2019, joblessness fell lower than many economists believed it could go, and inflation stayed below the Fed’s 2% goal, a fact that perplexed policymakers.

But the absence of inflation concerns allowed the Fed to respond to Trump’s tariffs. After raising rates to 2.5% in 2018, Fed officials held them there in the first half of 2019 before starting cuts in July.

“Ultimately, the risk of a slowing in the economy and a tick up in unemployment outweighed the risk of easing too much and creating excesses,” then-Dallas Fed President Robert Kaplan said in a recent interview. “That’s what won out.”

Former Chicago Fed President Charles Evans recalls what was almost a recession in manufacturing unfolding. “You could see it slowing down a part of the economy after taxes had been cut,” he said in an interview last week. “That was very surprising, I thought.”

Today, the picture is almost flipped. The US is coming off two years of strong growth and the highest inflation in 40 years. While the Fed’s tightening in 2022 and 2023 helped cool price growth, it still hasn’t reached officials’ 2% target.

That’s left the Fed hesitant to cut rates further after lowering them by a percentage point at the end of 2024. Policymakers held rates steady at their meeting concluding Wednesday, and Powell has made clear any further reductions depend on inflation continuing to fall — or a marked deterioration in the labor market.

Strategist Who Coined the Magnificent Seven Warns US Tech Is Set to Lag Sees risks from overexposure to US stocks as AI spending peaks

(…) Those positions are at risk as spending on artificial intelligence is set to peak, Hartnett wrote in a note. He also sees other catalysts of a US outperformance including excess fiscal support and immigration fading this year.

“US exceptionalism now exceptionally expensive, exceptionally well-owned,” the strategist wrote. “‘Magnificent 7’ becomes ‘Lagnificent 7,’ supports broadening of US and global equity and credit markets.” (…)

• Narrow markets are the exception, not the rule

2023/24 was the narrowest stock market since 1998/99’s technology bubble, which in turn was the narrowest market since the Nifty 50 period
of the early 1970s. With only three such periods in the past 50 years, narrow markets are the exception and not the rule.

Chart 1 shows how unusual 2023/24’s market has been. Just 30% of the constituents of the S&P 500® outperformed the index in 2023 and slightly less in 2024. That is similar to the 1998/99 period, and both are well below the long-term median of 48%. (…)

Extremely narrow leadership implies a scarcity of growth opportunities, i.e., there are very few companies that can grow. An extended period of narrow leadership, therefore, implies a terrible secular period during which the economy experiences an extended contraction/recession and the resulting scarcity of earnings growth forces companies to lay off workers and decrease capital spending. (…)

Unlike during the Great Depression, fundamentals haven’t justified today’s extraordinarily narrow leadership. 2023/24’s economy was healthy and late 2022/early 2023’s profits recession was quite mild compared to that during a depression.

If the recent market narrowness assumes that competition and innovation no longer matter, it may only be a matter of time before they broaden. When markets broaden from unusually narrow periods, however, they do so suddenly and for years.

The Goldman analysis also highlights that market volatility tends to significantly increase during the year after extremely narrow markets. Diversification typically constrains portfolio volatility, but narrow markets skew indices’ constituent weightings toward fewer companies effectively making an index less diversified. Increased concentration towards a small number of cyclical companies makes the overall market more susceptible to company-specific risks.

Because narrow markets are the exception and not the rule, history shows that markets broaden for years after periods of extremely narrow market leadership. Investors, however, tend to wait for the abnormally narrow leadership to reassert itself rather than repositioning portfolios for a more normal market.

This was certainly the case after 1998/1999’s narrow markets. Chart 2 shows S&P 500® sector performance during the final year of the Technology Bubble (3/31/99 – 3/31/00). Only 1 sector, Technology, outperformed the market during this period. (…)

The Technology sector quickly started to underperform and didn’t recover even after 5 years.

Diversification was critical to surviving the post-bubble period. Although the narrow leadership dragged down the overall market performance, many sectors posted significant positive returns.

  

Investors seem unduly confident despite the increased potential for volatility after narrow markets. Diversification today is considered a hinderance to performance by many investors rather than a risk reduction tool or one that might shift portfolios toward under-appreciated investments.

Chart 7 highlights the question in the Conference Board’s Consumer Confidence Survey that asks whether respondents think the stock market will be higher in 12 months. The past two months’ readings show a level of confidence in the stock market never seen in the survey’s history.

Investors’ willingness to take risk is even more startling. Chart 8 shows individual investors’ aggregate portfolio equity beta. Admittedly, we have re-published this chart from BofA Securities for many months. We thought it was extraordinary when their equity beta rose to 1.2, but it has continued to increase as the market continued to narrow. It is now an absolutely mind-boggling 1.7!

The juxtaposition of investors’ historic confidence relative to the performance history when markets broaden from extremes suggests that investors might be ill-prepared for volatility.

It appears that 2025 could be the year when investors’ extreme confidence and willingness to take significant portfolio risk comes face-to-face with the volatility associated with the broadening of a speculative market.

Diversification is one of the basic tenets of investing, but investors’ greed during speculative periods leads them to shun diversification because it looks stodgy, boring, and a lead weight on the potential for high returns. Themes regarding “The Magnificent 7”, “US exceptionalism”, and others reflect the current shunning of diversification. (…)

Today’s narrowness is even more dangerous than previous ones given the rising share of passive investments. A high beta works both ways…

(Apollo Management)

Goldman Sachs:

The equity markets’ correction, triggered by news of the DeepSeek LLM model, has been the first fall of more than 3.5% of the Magnificent 7 since last Autumn. In our view this is a correction and not the start of a sustained bear market.

Most bear markets are triggered by expectations of falling profits driven by fears of recession. Our economists remain confident about world growth and remain above consensus on their forecasts for the US, putting the probability of recession in the next 12 months at 15%.

We also expect interest rates to be cut, albeit modestly this year, alongside more progress on inflation moderation – and a cheaper entrant into the AI space might increase confidence in this trend. This combination of macro conditions has historically been supportive of risk assets.

Nevertheless, we have argued that context is important and that equity markets came into the year priced for perfection, leaving them vulnerable to disappointments. Returns in equity markets, led by the US, have been unusually strong over the past couple of years, and particularly since October 2023 as investors began to become more optimistic about the prospects for lower inflation and interest rates alongside a soft landing.

Overall levels of valuation had reached very high levels, particularly in the US, and P/E multiples have increased meaningfully since Q4 2023. In the case of the US, while much of this reflects the largest technology companies, the equity market remains expensive relative to history even if we exclude large cap technology (the second column). Furthermore, while other equity markets around the world are much cheaper than the US, most are not particularly cheap relative to their own history. The main exception is China.

While surging equity returns coupled with high valuation provide fertile conditions for a correction, it is also the concentration of equities as an asset class that has left equity investors vulnerable to disappointments. The rising concentration in equities has taken three forms: the growing dominance of the US equity market in the global index, the ascent of the technology sector, and the rise in single stock concentration (particularly in the US).

To be clear, these factors have emerged as a function of strong fundamentals, not as the result of speculation or irrational exuberance. The growing dominance of the US equity market has simply mirrored its relative profit growth since the financial crisis.

Meanwhile, the growing influence of technology on market returns reflects the significant outpacing of technology profits relative to other industries over the same period.

Furthermore, the dominance of the biggest companies in the US equity market is a function of their vastly superior earnings power over the past decade.

The dominance of US equity market, technology sector, and dominant companies does not represent a bubble based on irrational exuberance but is rather a reflection of superior fundamentals. Nevertheless, while stronger growth has justified the patterns of returns that have driven equity markets over the recent past, they do not tell us about the future expectations or returns.

On a forward-looking basis, while the US equity market is likely to move higher based on solid earnings growth, its superior earnings growth at the index level is set to fade, opening up opportunities for more diversification. For this year at the index level our top-down forecasts are similar across all regions (except for Europe), suggesting more opportunity for geographic diversification.

At the same time, our US strategists expect the strong pace of earnings growth of the dominant technology companies to continue but also to moderate on a relative basis.

The dominant companies are still expected to see superior growth but the gap is narrowing. We have argued that this leaves a risk in such a concentrated market. The price action of the GRANOLAS in Europe last year, provides a good use-case for the risk of high concentration in the US with the Magnificent 7. Despite their earnings continuing to grow faster than the rest of the European market, the GRANOLAS valuations have significantly declined, resulting in flat year-over-year performance. This derating has negatively impacted the overall market due to the size of these companies. (…)

History provides some useful lessons. First, the original capex spenders on revolutionary technology are not always the biggest beneficiaries; the experience of the Telecom companies in the late 1990s is a good example. Second, even very dominant companies eventually succumb to competition – often from new companies in the same sector – just as AMD and Intel experienced, for example, with the ascent of Nvidia.

The extent to which these observations are relevant to the current market setup is still not clear. But the news around DeepSeek has been a wake up call that has shaken the confidence that was reflected in market pricing. Indeed, our technology analysts argue “DeepSeek has introduced pricing competition into the foundational model layer at a point in time where models are just about good enough for many enterprise use cases”. The revelation of a cheaper competitor entering the AI space has exposed the risk of concentration. This is relevant because equity markets are not typically driven by absolute outcomes, but rather by outcomes relative to expectations.

We noted last week that the polling responses to questions asked at our annual Global Strategy Conferences revealed a very strong consensus that the patterns of leadership that drove the returns over the past couple of years is set to continue, leaving room for disappointment. For example, at our London conference, 58% of the votes – the highest ever – were for the US equity market to be the best performer in 2025.

This marks an increase from 32% last year and 18% the year before. Meanwhile, Europe remains the least preferred, receiving only 8% of the votes compared to 13% last year. We find similar examples of US exceptionalism as a driver of outperformance in the responses at our Asia Strategy conference where a record low 3% believed Europe would outperform.

There was a similar degree of confidence in the ongoing dominance of the technology sector as the likely winner this year. (…)

While the macro set up for equities remains supportive, we have been recommending that investors stay invested but diversify to improve risk adjusted returns.

The line to watch on this chart id the yellow Rule of 20 Fair Value line which is EPS x 20 minus inflation. Profits are still rising faster than inflation.

Fed Stands Pat on Rates, Entering New Wait-and-See Phase Chair Jerome Powell says ‘We do not need to be in a hurry to adjust our policy stance’

(…) With interest rates now “significantly less restrictive” than they were before last year’s cuts, “we do not need to be in a hurry to adjust our policy stance,” said Fed Chair Jerome Powell at a news conference after the meeting.

Those comments suggested the central bank was likely to stay on hold at its next meeting in mid-March. Powell said the Fed would need to see “real progress on inflation” or unexpected weakness in the labor market before considering further rate reductions. (…)

“We seem to be set up for further progress” on inflation, he said. Being set up for progress is one thing, “but having it is another.” (…)

Minutes from the Fed’s previous meeting, released earlier this month, said a substantial majority of officials still judged their policy stance at the time to be “meaningfully restrictive.” (…)

John Authers:

Everyone knew that the Fed wouldn’t change interest rates. They also did nothing to change their policy of gradually reducing the balance sheet. Instead, the greatest excitement in its communique came from a slight alteration in language about inflation and unemployment. Last month, the FOMC statement said:

Labor market conditions have generally eased, and the unemployment rate has moved up but remains low. Inflation has made progress toward the Committee’s 2 percent objective but remains somewhat elevated.

This month that changed to:

The unemployment rate has stabilized at a low level in recent months, and labor market conditions remain solid. Inflation remains somewhat elevated.

So if the labor market is solid rather than easing, and the Fed no longer says it’s making progress toward the 2% target, that sounds like a distinct shift toward hawkishness — a propensity to make the next move a hike. Then Powell was asked about what appeared to be a critical signal in his Q&A. His response:

If you just look at the first paragraph, we did a little language cleanup there… We just chose to shorten that sentence.

(…) The effect of the Fed’s language cleanup — and Powell’s subsequent attempt to clean up how people had interpreted it — was clearly visible across broader financial markets. The dollar, the gold price, the 10-year Treasury yield, and the S&P 500 all moved sharply. Those moves were completely reversed on Powell’s clarification. In the final analysis, the Fed did a good job of convincing people that it wanted to cut rates, but was in no hurry to do so.

Bank of Canada Cuts Rates, Warns of Economic Shock From Trade Conflict Policy rate down quarter-point, officials warn of tricky balancing act with potential of weaker growth, higher prices

Bank of Canada Gov. Tiff Macklem said the central bank would be limited in offsetting the damage from a trade row, because policymakers would have to wrestle with both weaker growth and higher prices. The central bank sets interest rates to achieve and maintain 2% inflation.

The Bank of Canada can “try to be a source of stability” through a potential trade disruption, Macklem said, “so the adjustment is less unpleasant than it otherwise would be.” He added, “We can’t lean against weaker output and higher inflation at the same time.”

Canada’s central bank lowered its policy rate by a quarter-point to 3%, as widely expected in a survey of economists last week by The Wall Street Journal. In its decision, the Bank of Canada said a recent rebound, fueled by aggressive rate cuts since last June, is in jeopardy should President Trump follow through on his threat of a 25% tariff on imports from Canada and Mexico. Canadian officials have vowed to retaliate forcefully with their own tariffs in such an event.

“A long-lasting and broad-based trade conflict would badly hurt economic activity in Canada,” Macklem said. (…)

Roughly three quarters of all Canadian exports are U.S.-bound, and Trump has said the hefty tariffs could begin as early as Saturday.

Initially, Trump said the 25% tariff would be in place until Canada and Mexico fortified their border security. He has since expanded his reasoning, complaining about a trade deficit Canada runs with the U.S. (…)

The central bank, in its decision Wednesday, removed any guidance about the future direction of interest rates. At his press conference, Macklem outlined scenarios in which officials might either have to cut rates to support growth, or raise them to prevent one-time cost increases from tariffs from feeding into wages and other prices. (…)

Bank of Canada analysts modeled a scenario in which the U.S. imposed a 25% tariff on all imports, and its trading partners responded in kind, over a multiyear period. In that model, the trade row would knock down Canadian growth by 2.5 percentage points in the first year. As a result, the 1.8% growth forecast for Canada this year would turn into a decline of 0.8%, or a recession. (…)

The central bank published a quarterly forecast. Absent tariffs, the Bank of Canada expects 1.8% growth in both 2025 and 2026, or a downgrade from the previous forecast of 2.1% and 2.3% respectively. Officials attribute the downward revision to a projected slowdown in population growth stemming from immigration-policy changes.

Hiring has picked up, although Macklem said the labor market remains soft and there are indications wage growth is easing.

AI CORNER

Microsoft, Meta Talk Up Their Big AI Ambitions and Spending Plans Tech giants vow to barrel ahead despite jolt from China’s DeepSeek

(…) Microsoft Chief Executive Satya Nadella and Meta CEO Mark Zuckerberg said in earnings calls that DeepSeek had made real innovations. But the two portrayed the Chinese company’s work as part of a technological evolution that will make AI cheaper and more widely used rather than the extensive disruption that some observers have perceived.

“There’s a number of novel things that they did that I think we’re still digesting,” Zuckerberg said on a call with analysts after Meta reported record quarterly sales and strong profit growth. “I continue to think that investing very heavily in [capital expenditures and infrastructure] is going to be a strategic advantage over time. It’s possible that we’ll learn otherwise at some point. But I just think it’s way too early to call.” (…)

“In some sense what’s happening with AI it’s no different than what was happening with the regular compute cycle. It’s always about bending the [cost] curve,” Nadella said. (…)

Wednesday he said simply that “as AI becomes more efficient and accessible, we will see exponentially more demand.” (…)

Zuckerberg on Wednesday emphasized that the winners of the AI arms race—both countries and companies—will be able to set the terms.

“There’s going to be an open-source standard globally,” he said on Meta’s earnings call in response to a question about DeepSeek. “For our, kind of, own national advantage, it’s important that it’s an American standard,” he said. “So we take that seriously and we want to build the AI system that people around the world are using.”

Zuckerberg said DeepSeek did “a number of novel things” that Meta hopes to implement in its systems. (…)

He predicted that companies investing vast sums in computing infrastructure will be able to deploy it in ways that can adjust as AI advances, and such investments will continue to be an advantage. More computing power can generate a higher level of intelligence and service, he said. (…)

“People don’t all want to use the same AI,” he said. “People want their AI to be personalized to their context, their interests, their personality, their culture, how they think about the world.”

Microsoft said it expects the next two quarters to be similar to the $22.6 billion it spent in the last quarter, which is roughly in line with Wall Street’s expectations. The company didn’t say if DeepSeek’s breakthroughs would eventually lead to a big reduction in spending on AI computing—the key fear driving Nvidia’s selloff.

• U.S. Businesses Already Love DeepSeek The Chinese company’s new model promises to lower the cost of AI for enterprises—even amid concerns about cybersecurity and geopolitics

(…) New York Life Chief Data and Analytics Officer Don Vu said the insurance company is exploring the use of DeepSeek’s AI model. Vu said the company has a framework for evaluating the effectiveness of different models, like OpenAI’s GPT, Anthropic’s Claude and Meta Platforms’ Llama, for different use cases. It will now test how effective DeepSeek’s new model is in areas such as service and claims.

What New York Life won’t be doing, Vu said, is using the existing DeepSeek app, which raises data security questions. It will instead download the open-source version and begin experimenting with that.

For security reasons as well as ease of use, some enterprises will prefer not to host their own versions of the DeepSeek model and will turn to vendors like Amazon that make them available through their platforms. Smaller versions of new DeepSeek models are now available on Amazon Bedrock, the company’s AI development platform, for enterprises to test, run and use for business cases, a company spokesperson said.

German software company SAP said it is open to leveraging AI models coming from Chinese companies like DeepSeek if they meet certain cost, reliability and data privacy requirements, its Chief Financial Officer Dominik Asam told The Wall Street Journal. (…)

The Chinese ownership is a clear deterrent for some CIOs who might otherwise consider leveraging the tech, even as they remain optimistic about the pressures it could put on U.S. tech companies.

Marc Kermisch, chief technology officer of Emergent Software, said tech leaders should consider blocking the app for overzealous employees who might inadvertently put corporate data into it, although he said he is experimenting with it in a controlled environment. Given its Chinese ownership, Kermisch said he is skeptical that DeepSeek’s model will become viable for widespread use by American companies. But he said he is still hoping it can take a role in pressuring U.S. model makers to drive down costs. (…)

Brian Greenberg, CIO of leadership consulting firm RHR International, said he considers DeepSeek’s new model “fascinating and worthy of exploration.” However, that excitement should be tempered. To use the open-source version of DeepSeek inside the company, Greenberg added, would require a strict cybersecurity review.

Adnan Masood, chief AI architect of digital technology and information-technology services firm UST, said, “The anxiety is that you’re feeding sensitive corporate data into a system that originated from a strategic adversary, no matter how ingenious the engineering.”

“On the flip side, drastically lower costs and advanced capabilities tempt executives to risk these concerns for competitive advantage,” he added.

• It’s Not Just DeepSeek. A Guide to the Chinese AI Companies You Need to Know. Chinese internet giants and startups are competing with each other and U.S. rivals

Alibaba: The e-commerce giant provides conversational chatbot service Qwen, powered by multiple AI models, including some designed for more complex reasoning and coding tasks. This week, it also released Qwen2.5-Max, an artificial-intelligence model it said was competitive with global leaders, including DeepSeek. It hasn’t clarified whether it developed the model with the low cost and high efficiency that DeepSeek has boasted of.

Tencent: China’s biggest videogame company has developed multiple versions of its AI model Hunyuan. It said one version released in November delivered performance comparable to Meta’s Llama 3.1. According to some researchers, Tencent might use around a tenth of the computing power Meta used to train the model. The company is integrating AI capabilities into its WeChat app, the ubiquitous platform in China providing everything from chats to banking.

Baidu: Baidu, which first emerged as a search-engine company, was the first in China to launch a ChatGPT equivalent, called Ernie Bot. Its technology chief said in November that its model had 430 million users.

ByteDance: The owner of TikTok has a chatbot called Doubao, which can be translated as bean bun. The app has been among the most downloaded chatbots in China, with around 60 million monthly active users, according to Aicpb.com, a website tracking AI products.

DeepSeek: The company surprised the global tech community earlier this month after it said it had trained AI models that delivered high performance at low cost and without the most advanced chips. Then, on Tuesday, it released a multimodal model, called Janus Pro, that it said could produce results comparable to OpenAI’s text-to-image model DALL·E 3.

StepFun: The company, valued at around $2 billion, has a model that is now ranked for performance among the top 10 in the world in Chatbot Arena. Founded by a former senior Microsoft scientist, the company counts Tencent and the Shanghai government as key investors.

Moonshot AI: Moonshot’s Kimi chatbot has around 13 million users in China, according to Aicpb.com. The startup, valued at around $3.3 billion and backed by Alibaba and Tencent, was founded by a young Chinese scientist who had stints at Meta and Google. This month, Moonshot released a multimodal reasoning model, called k1.5, that it said outperformed big names such as OpenAI’s GPT-4o and Anthropic’s Claude3.5 Sonnet on some major benchmarks, including a math challenge.

MiniMax: MiniMax is a Shanghai-based startup valued at $3 billion. It invented a Character.ai-like companion chatbot called Talkie, which has become popular in the U.S. This month, it published two open-source models that it claimed to be comparable to OpenAI’s GPT-4o and Anthropic’s Claude3.5 Sonnet, using a technique called Lightning Attention that allows faster computation.

Zhipu: Zhipu, valued at around $3 billion in its latest fundraising round in December, has invented a chatbot, as well as a video-generating model called Ying that is similar to OpenAI’s Sora. Zhipu was also included this month in a U.S. trade blacklist for developing AI systems that could have military uses. Zhipu said the U.S. move was baseless.

• DeepSeek’s ‘Open AI’ Should Terrify Sam Altman It’s not the only company building systems that are free to use.

(…) Across the world, similar open-source models are being built, ready to eat into Altman’s market share with systems that are vastly cheaper or even free to use. That’s an unnerving prospect for businesses like OpenAI and Anthropic, whose primary path to profitability is selling access to a foundation model for a higher price. (Alphabet Inc.’s Google and Microsoft Corp. at least have cloud and software businesses to fall back on.) Publicly available models like DeepSeek might not reign supreme in the end, but they could win an uncomfortably large share of the pie, and fulfill promises1 that Altman has neglected around openness and collaboration.

“DeepSeek’s paper on [its new AI model known as] R1 was more transparent than anything I’ve seen from OpenAI since GPT-3,” says Gary Marcus, a professor emeritus at New York University, who has long bemoaned OpenAI’s opacity and warned of an AI market crash.

That openness is part of a shift happening in China. (…) The Chinese company has put most of the details about R1 on the internet, something that OpenAI and Google wouldn’t do today. Its breakthrough is essentially a novel approach to so-called reward modeling, according to one summary of the findings.

(…) DeepSeek isn’t technically “open source,” according to the Open Source Initiative (OSI), a global authority that certifies software licenses. “They do not provide the source code to process and filter and train the data, nor the information about the training data,” a spokesman for the OSI tells me. “Others would have to build this from scratch based solely on the paper they published.”

But as a so-called open-weight system, DeepSeek can still provide the crucial model parameters that others can copy and improve upon, unlike OpenAI, which keeps its models locked in a black box. (…)

China’s shift toward building open-source models has come about in the last few years, driven by US export controls, government support and the impact of Meta’s own open-weights model, Llama, which several Chinese tech firms have been enthusiastically fine-tuning for their own applications, according to Computer Weekly.

For the entrenched giants of Silicon Valley, the real threat isn’t just China, but the way Chinese tech firms could fuel even greater momentum for the open-source movement in other parts of the world — like France, home to promising open-source AI firms Mistral AI and Hugging Face Inc., and the US, where research collective EleutherAI has been influential in creating models like GPT-J and GPT-NeoX.

Of course, open-source AI has downsides, the obvious one being misuse by bad actors. And while DeepSeek’s new model has solved the problem of cost, it hasn’t solved hallucinations. Plenty of examples exist on social media of users jailbreaking R1 to get around its Chinese Communist Party-friendly filters. And while the model was cheap to build, it is still relatively expensive to run, Marcus points out.

Yet DeepSeek is still giving open-source AI newfound credibility among software builders. Developers on forums like X have been showing off all the apps they’re building on DeepSeek’s free platform. (…)

OpenAI’s business of building foundation models has become commoditized. The real economic potential is for those building products on top of all that infrastructure, or on top of platforms like DeepSeek’s that are free to use.

The irony is rich. When OpenAI launched in 2015, it promised to “freely collaborate” and share its patents with the world. Instead, in its galactically ambitious quest for human-level AI, it began morphing into exactly what it set out to disrupt: an opaque, money-hungry tech giant, which is now being challenged by genuinely open AI.

Altman convinced the world that building powerful AI required massive resources, but he’s now learning that his moat isn’t that deep after all. Perhaps the future belongs not to those who can spend the most, but to those who dare to share their workings with the world.

FYI: The Most Important Time in History Is Now AGI Is Coming Sooner Due to o3, DeepSeek, and Other Cutting-Edge AI Developments