news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

reporter: wu yifan

editor: wu yangyang

Key Points

all qualcomm wants is intel's chip design business, not chip manufacturing;

this deal could become the largest deal in global technology history, but it could also fail to close due to antitrust scrutiny.

since the 2010s, the revenue from cpus, which is intel's core business, has hardly grown, especially for personal computers (i.e. pcs);

intel's foundry business also doesn't look as healthy as the outside world thinks, and intel seems to have been hiding this from its investors through financial means in the past;

amd's former ceo said that for the past 30 years, intel has been focused on suppressing amd, which has caused them to ignore important trends in mobility and low power consumption.

on september 21, it was reported that qualcomm had finally taken the step of approaching intel to discuss a tender offer for the latter. if the acquisition is successful, it will become the largest deal in the history of global technology, surpassing microsoft's acquisition of activision blizzard for $75 billion in 2023. on the last trading day before the news was exposed, intel's market value was $93 billion.

qualcomm has been discussing the acquisition for several months, and on september 5, it was reported that it was interested in acquiring intel's chip design business, not chip manufacturing. at the time, intel insisted that the company was "firmly committed to the personal computer (pc) business" in response to the rumors. after all, pc processors have been the core of intel since the 1980s and the company's business that has contributed the most revenue and profits to date. at that time, qualcomm had not yet contacted intel about potential acquisitions.

the news on september 21 made the acquisition more realistic. at present, intel's market value is not high, less than half of qualcomm's and less than 1/30 of nvidia's. moreover, intel's equity is quite dispersed, with its largest shareholder, pioneer, holding only 9.2% of the shares, and the top ten shareholders holding more than 30% of the shares combined. most importantly, these shareholders are almost all wealth management companies, all financial investors.

due to antitrust factors, there is still a high possibility that the deal will not be completed. in october 2016, qualcomm tried to acquire nxp semiconductors, the world's largest automotive chip provider, for about $38 billion. however, until the deadline for the transaction, the deal was not approved by chinese regulators. if the acquisition of intel is successful, qualcomm will have a competitive advantage in both smartphones and personal computers.

qualcomm has already tried to enter the personal computer field with the help of generative artificial intelligence. at the computex conference held in taipei, china in june this year, qualcomm released the snapdragon x series aipc chip based on the arm architecture. on september 4, it released a new aipc chip "snapdragon x plus 8-core", targeting the pc sinking market of us$700 to us$900. together with the "snapdragon x elite" and "snapdragon x plus 10-core" that have been launched, even if the acquisition of intel fails, qualcomm's investment and products in the pc processor field can already support multiple market segments from high-end to mid-to-low-end. according to the plan, qualcomm will also increase investment in the pc chip field, trying to increase the shipment volume of its pc business to three times that of last year.

there is still a lot of uncertainty in the deal between qualcomm and intel, but the difficulties intel faces are set in stone.

on august 7, less than a week after intel released its second-quarter financial report, its shareholders collectively sued it in the u.s. federal court in san francisco, claiming that "investors did not know that intel's foundry business was in trouble and that its losses were billions of dollars more than investors thought."

the matter originated from an adjustment to intel's financial reporting rules in april this year. at that time, intel independently accounted for all foundry businesses, including internal and external foundry, in its financial reports for the first time. previously, the foundry business in intel's financial reports had never reflected the level of profit when it was doing foundry for internal orders. simply put, before the adjustment, the financial report showed that intel's foundry business did not lose much, with a loss of only us$482 million in 2023; but after the adjustment, the loss was as high as us$4.229 billion in the first half of 2023 alone. in other words, intel's foundry business does not look as healthy as the outside world imagines, and intel seems to have concealed this from its investors through financial means in the past, which angered investors.

the adjustment of financial rules is most likely intel's financial preparation for the split of its foundry business. in june 2023, intel announced the split of its foundry business, completely separating design and manufacturing to avoid competition with foundry customers.

before this, intel's foundry business was not large and was far from comparable to tsmc. according to the old financial standards, in fiscal 2023, intel's foundry business revenue was us$952 million, which was 1/72 of tsmc's revenue of us$69.298 billion.

after the revision of the financial reporting rules, intel's annual revenue from foundry services can reach about $18 billion, which is about 1/4 of tsmc's revenue. however, the loss will quickly climb to $8 billion, which means that when intel's foundry business is a huge loss-making business for upstream design companies of the same size as itself.

of course, intel’s modification of the financial report disclosure rules can be understood as just a financial measure. its foundry business may not be that bad. intel’s real purpose may be to continue to cover up the sluggish sales of its main business - chips (whether cpu or gpu) - by sacrificing the profit and loss statement of its foundry business.

missing the mobile wave

since the 2010s, the revenue of cpu, which is intel's core business, has almost stopped growing, especially the cpu used in personal computers (pcs). in 2015, intel's client computing division had revenue of $32.2 billion. by 2021, this part of revenue only grew to $40.5 billion, with an average annual growth rate of less than 10%. in some years, it only grew by 0.3%, and it fell continuously in 2022 and 2023. in 2023, intel's revenue from selling cpus in the pc market will only be $29.3 billion, 11% less than in 2015.

this makes sense because since 2010, the consumer electronic devices that more consumers have bought home more frequently are not personal computers, but smartphones - the former is a position that intel has pioneered and dominated, but the latter is not.

as to why intel missed the wave of mobile internet, a widely circulated view is that intel's then ceo paul otellini rejected apple's order in 2006. that year, apple founder steve jobs approached otellini and proposed a demanding order in his usual bargaining style: he wanted intel to develop a new chip for the upcoming iphone, but the bid was far lower than intel's estimated cost of producing such a chip.

intel did not accept the order given by jobs. in 2007, apple released the first generation of iphone, which was manufactured by samsung. in 2009, samsung's market value reached $110.2 billion in one day in february, $860 million higher than intel, making it the world's most valuable semiconductor manufacturer. in november 2012, otellini suddenly announced that he would resign as ceo in may 2013 without confirming the new ceo candidate. the outside world speculated that it was related to his misjudgment of the mobile internet wave.

otellini himself also regretted his misjudgment. "in my career, i often made decisions based on my intuition. i should have followed my intuition. my intuition told me to say yes to the iphone. but we all like to speak with data," otellini said in an interview with the atlantic in 2013. the numbers told otellini and intel that the processors developed by intel for personal computers could be sold for $100, while the processors for mobile phones might only cost $10. moreover, according to estimates, the fees given by jobs could not cover intel's production costs, even if the order volume was larger. later facts showed that intel had both overestimated the costs and greatly underestimated the scale of equipment in the mobile market.

if otellini had accepted the iphone order at that time, intel might be very different today, but it would not necessarily change intel's fate in the mobile internet era, because the real reasons why intel missed the mobile market were several other things: x86 architecture, amd, and profit margins.

x86 is a chip architecture invented by intel in its early days, with the characteristics of high performance and high power consumption. on the pc side, high power consumption is not a problem, the key is performance. but for mobile devices with smaller size and battery capacity, the importance of power consumption and performance is reversed compared to the pc side. in the mid-2000s, mobile device manufacturers began to collectively turn to a standard called arm. the chip performance of this architecture is far inferior to intel's x86 architecture, but the power consumption is much lower.

intel is not ignorant of technology trends. as early as the 1990s, when grove was still ceo, mobile devices were a topic of frequent discussion within intel. at a headquarters meeting in the early 1990s, an intel executive waved his handheld computer and declared that "these devices will grow up and replace personal computers." in 1997, when dec declared bankruptcy, intel actively acquired its strongarm team. as the name of the team suggests, the strongarm chip it developed is also based on the low-power arm architecture, and its performance is stronger than the public version of the arm architecture. around 2000, intel began to launch the stongarm-based processor - xscale.

but in 2006, the same year that otellini rejected the iphone chip order, intel sold xscale to communications chip company marvell technology for $600 million. compared with the unknown outbreak of mobile devices, intel's more urgent matter was to deal with competition from amd.

in the early 2000s, as the main supplier of pc cpu products, intel's market share was higher than amd, but the gap was not large, and the market share of both companies hovered around 50%. in 2003, amd took the lead in entering the 64-bit processor market with a 64-bit cpu. intel's processors at the time were only 32 bits, and users only needed to look at the numbers to know which one was more advanced. like smartphones, the cpu market also updates products every year according to moore's law (the number of transistors integrated on a chip of the same area doubles every 18 months), and the advantages or mistakes of one generation of products may reverse the market structure.

with the 64-bit cpu launched one step ahead, amd almost caught up with intel in the desktop processor market in 2004. in the first quarter of 2006, amd even surpassed intel. as of that quarter, intel's sales had not met wall street's expectations for two consecutive quarters, and its operating income in 2006 only increased by 5% compared with 2000. the board of directors asked otellini to explain why the company's performance declined so severely and accused the company of being too bloated. otellini quickly got rid of xscale, and it took another two years until 2008 to truly destroy amd's momentum and enable intel to regain its leadership in the pc field.

after selling xscale and ending its brief battle with amd, intel did not completely give up its layout in the mobile field. in april 2008, intel launched the atom series of chips for mobile devices, a project that intel had laid out as early as 2004. unlike the xscale chip based on the arm architecture, the atom series of chips is based on intel's traditional x86 architecture - the sale of xscale is considered to be a choice made by intel between arm and x86.

the instructions given to the atom development team were clear and firm: low power consumption, based on x86, and backward compatible with intel's existing products. the r&d team developed atom according to the instructions, but this product line never received enough resources before the launch of iphone 4. when atom was first launched in 2008, it only supported nettops and netbooks. the official chip version for smartphones and tablets was launched in 2012.

the smartphone market was changed forever in 2011. in 2009, apple's shipments were less than one-third of nokia's, which was the mobile phone brand with the highest market share that year. in 2010, apple released the iphone 4, and in 2011 it surpassed nokia. by 2013, nokia had almost disappeared from the mobile phone market. intel was forced to announce the cessation of the development of the atom series of processors in 2016, withdrawing from the mobile phone and tablet computer fields, and shifting the company's focus to data centers and foundry.

the reasons why intel rejected the iphone processor order, sold xscale, and failed in atom were different, but they were similar in one way - internal resistance based on profit margins.

it has been reported that some intel executives realized the importance of mobile technology to the company's long-term development and wanted to vigorously promote atom, but were hindered because another group of people were worried that the atom processor and the low-cost devices it supported would have an impact on the company's traditional business. therefore, they were unwilling to transfer production and design resources from the pc and server chip departments - which sounds like the same considerations as when rejecting iphone orders and selling xscale.

the obstacle also comes from the manufacturing link. intel's wafer fabs, manufacturing strategies and resources are all oriented to expensive large desktop processors, rather than low-cost mobile chips. if intel prioritizes atom, it will need to reorganize at least some wafer fabs and reduce costs in order to compete with arm processors produced by samsung and tsmc. in addition, if it fully invests in the mobile chip market, users may buy cheap mobile chips to put in their notebooks, weakening the demand for desktop chips with higher profits, which will affect intel's revenue.

history always repeats itself

whether intel, which failed to gain a foothold in the mobile internet era, can catch up with the ai ​​wave and successfully break through has become a problem that this veteran chip giant has to face next. however, intel has lost its voice in the gpu market, which symbolizes the era of artificial intelligence. in 2023, nvidia and amd accounted for 95% of the data center gpu market, and intel's gpu was nowhere to be seen.

the factors that caused intel to miss the opportunity in the era of artificial intelligence are still within that range: underestimation of emerging markets, repeated and unclear strategies, and obsessive adherence to the x86 architecture.

intel has always been worried about being left behind by the trend of an increasingly visual world. as early as 1998, the cpu giant jointly launched a gpu with gpu manufacturer real3d to enhance the visual effects of 3d games and dvds on personal computers. in 2005, the second year after the launch of the atom mobile chip project, a gpu project called larrabee was also established. however, like atom, larrabee once again chose the x86 architecture, and it is not a stand-alone graphics card that focuses on pursuing top-level graphics computing effects like nvidia's gpu, but a hybrid chip combined with intel's existing cpu. this approach is intended to enhance intel's old business-the capabilities of the cpu itself, rather than seeking a share of the stand-alone gpu market.

in 2005, gpus did not seem to have many more use cases besides graphics computing. but at that time, nvidia had realized that the high-speed parallel computing of gpus could not only be used to render graphics, but also in more computing fields beyond that - such as artificial intelligence. so in 2006, nvidia began to invest heavily in a software called unified computing device architecture (cuda), allowing programmers, not just graphics experts, to use nvidia chips. larrabee was soon cut off by intel after its first attempt at graphics computing was not successful. pat gelsinger, then cto and head of the larrabee project, left the company in 2009 in a rage.

in the seven or eight years since intel withdrew from the gpu market, the field of artificial intelligence has experienced a wave of computing dominated by alphago and image recognition. faced with the gpu demand that can no longer be ignored, in 2018, intel announced its return to the gpu market and established the accelerated computing and graphics division (axg) for the first time in the first quarter of that year. it not only poached raja koduri, who was in charge of the graphics department from amd, but also invited back tom forsyth, the father of larrabee. it looks like it is ready to make a big move - but it only looks like it.

in 2021, intel launched an independent graphics card brand called arc, planning to ship more than 4 million independent gpus in 2022, and expecting this business to contribute more than $10 billion in revenue by 2026. however, the insufficient performance of arc products has prevented this series of plans from being fulfilled - its architecture can complete lightweight tasks well, but it is not suitable for high-performance tasks such as data centers. in may 2022, the ai ​​processor gaudi 2 released by the habana labs team acquired by intel had a similar fate to arc. gaudi 2 is said to be suitable for both training and reasoning, but its speed is not as fast as nvidia's popular h100 gpu.

this series of results is not surprising. compared with nvidia and amd, intel has invested too little time in gpu research and development. the resources invested are also insufficient. nvidia's investment in cuda software alone exceeds $10 billion. in comparison, intel's total investment in gpu development is only about $3.5 billion.

at the end of 2022, intel will disband axg, which was established less than two years ago. the graphics team will join the client computing group (ccg) and be combined with intel's traditional main business cpu, and the accelerated computing team will be incorporated into the data center and artificial intelligence business group (dcai). this split means that intel realizes that gpus for consumer devices and gpus for data centers are not the same product. the former can only solve lightweight tasks of running small models on the end side, while the latter needs to cope with various large model training tasks.

this architectural adjustment is correct but not timely enough. in the fourth quarter of 2022, intel's market share in the entire data center (including cpu and gpu) was 46.4%, but it fell to 19.1% in the third quarter of 2023. during the same period, nvidia's market share soared from 36.5% to 72.8%. accordingly, nvidia's series of gpu chips dedicated to data centers generated $18.4 billion in revenue in 2023, a year-on-year increase of more than 400%, while intel's data center revenue fell by more than 40% from its peak.

new technologies completely changed the market landscape within a year, a situation that intel had already experienced once in the smartphone era. coincidentally, even the terrible experience of rejecting the new era ticket was experienced again by intel 10 years later. it is reported that in the months of 2017 and 2018, executives of intel and openai discussed various options, including intel's acquisition of 15% of openai's shares for $1 billion in cash, and intel could acquire an additional 15% of openai's shares if intel could provide chips to openai at cost price. at that time, openai was very interested in getting intel's investment because it would reduce their dependence on nvidia chips. but intel eventually gave up, partly because then-ceo bob swan believed that "generative ai will not enter the market in the near future", and another possible reason was that intel's data center division did not want to produce products at cost price. do these two sentences sound familiar? i wonder if swan will express regret in a similar way to otellini when he has the opportunity to publicly recall his wrong decisions one day.

today, openai’s valuation has exceeded $80 billion, and nvidia, which supplies chips for training large models, once had a market value of more than $3 trillion, making it the world’s most valuable company.

it is worth adding that when intel started negotiating with openai in 2017, amd, intel's arch-enemy, suddenly became powerful again and launched the ryzen processor that year. ryzen means "zen" (rebirth) and "risen" (hope), with 8 cores and 16 threads, while the top intel core i7-6700k processor at that time only had 4 cores and 8 threads. it was also in 2017 that tsmc and samsung successively achieved mass production of 10nm chips, surpassing intel in chip technology for the first time. having lost the two major fortresses at the same time, intel once again had no time to take care of the era of artificial intelligence that was coming at an unknown time, until the market changed overnight.

in 2021, intel, which no longer wanted to miss any opportunities, invited kissinger back from outside to be the ceo. this is the first ceo in intel's history with a cto background. the two previous ceos, otellini and swan, who had rejected iphone and openai respectively, both majored in economics.

separate survival

kissinger, who took over intel again, has a series of new policies for this big ship that is almost difficult to turn around. the core is a strategy called "idm 2.0". idm stands for integrated device manufacture (vertical integration model), which means the integration of chip design and manufacturing. intel adopted this model when it was first established in the 1970s, designing and producing its own chips. however, since the 1980s, the fabless model and foundry model that separate chip design and chip manufacturing have been popular in the industry. nvidia, qualcomm, mediatek, etc. are all fabless companies, while tsmc is a typical foundry company. in fact, the reason why nvidia, qualcomm, and mediatek were able to successfully enter the market in a light-asset manner is precisely because of the existence of foundries such as tsmc.

intel is almost the last company in the industry that sticks to the idm model. even its arch-rival amd divested its manufacturing business in early 2009 and became an asset-light fabless company.

adherence to the idm model is believed to have dragged down intel on multiple levels. first of all, in terms of chip technology. as the originator of moore's law, intel has been proving this law with actions since the 1970s, until 2014 - that year, it took the lead in launching 14nm chips, when tsmc's most advanced process was still stuck at the 20nm stage. but then intel suddenly stalled. the 10nm chip, which was originally planned to be mass-produced in 2016, has not been fulfilled because the yield problem could not be solved. samsung and tsmc, which were originally lagging behind, took the lead in overcoming the yield problem of 10nm in 2017 and achieved mass production. the yield problem of intel's 10nm chip was not solved until the second half of 2019. until now, intel's chip process is still three generations behind tsmc. tsmc's 2nm chip has begun trial production in july this year, and intel's most advanced chip process has so far remained at 7nm, with 5nm and 3nm in between.

the yield issue seems to be a technical issue, but to a large extent it is also a business issue, as tsmc's two attempts to overtake intel in process technology illustrate.

historically, tsmc first caught up with intel in terms of process in 1999. at that time, the chip process of the entire industry was still at the micron level and had not entered the nano era. tsmc, which specializes in foundry business, received an order from nvidia, which hoped to use tsmc's production line to produce a processor code-named "geforce 256". this is the world's first true graphics processing unit (gpu). it was because of this order that tsmc broke through the 0.18um process for the first time and caught up with intel's newly released pentium iii 500e processor in terms of process. the second time tsmc caught up with or even surpassed intel was in 2017 with the 10nm process. at that time, tsmc received an order from apple for the a11 bionic chip, which was to be installed on the iphone 8.

tsmc's former chief technology officer, hu zhengming, believes that this is where the foundry model is superior to the idm model. by manufacturing chips for enough customers, tsmc has more opportunities to polish the process. "when you have a large number of foundry customers, the product cycles of these customers are not all synchronized. almost any time you have a new technology, there will be some customers willing to pay for it."

according to kissinger's media interview at the beginning of his tenure, one of the reasons why intel was stuck on the 10nm process for so long was that it delayed the adoption of euv lithography machines, trying to achieve 10nm on the less advanced duv through multiple exposures. by 2020, half of all euv lithography machines produced were installed at tsmc. in contrast, intel has just started using euv in its manufacturing process.

the new plan proposed by kissinger is nominally called idm 2.0, which is essentially to separate chip design and chip manufacturing, which is a relief for both teams. after splitting out the foundry business and establishing an independent foundry services department (ifs), this part of the business will be able to compete directly with tsmc and samsung; intel's own enemies will also be more single. from then on, it only needs to deal with chip design manufacturers such as amd, qualcomm, and nvidia, without having to divert its attention to dealing with tsmc and samsung. if necessary, in order to ensure that the product is launched in time, intel can even choose to outsource the chip. since the beginning of this year, intel has outsourced two chips code-named arrow lake and lunar lake to tsmc, using a 3nm process that intel does not have.

kissinger asked the ifs business to report directly to him, and he himself had great expectations for the foundry business. according to the 2025 roadmap announced in july 2021, intel plans to achieve a leap from 10nm to 2nm in these four years and regain its leading process position by 2025. kissinger himself predicted that ifs would break even by the end of 2030 and become the world's second largest foundry that year. according to data from international consulting firm counterpoint, in the fourth quarter of 2023, tsmc is still the first among the world's top ten wafer foundries, accounting for 61% of the market share; followed by samsung, accounting for 14%; intel is among the top ten, but its market share is less than 1%.

since 2021, intel has announced plans to build wafer fabs in arizona, ohio, germany, italy, ireland and other places, with a total investment of more than $100 billion. however, two years later, only the german wafer fab has given a clear start time and construction blueprint. the factories in ireland and poland are still in the subsidy negotiation stage, and the projects in italy and france are directly stranded. the latest news about intel's foundry business is that on september 16, intel announced plans to transform its chip foundry business unit into an independent subsidiary and allow it to obtain external financing.

the challenge facing kissinger is not only the gap between ifs's ambition and reality, but also that he needs to better sort out and promote intel's ai chip layout in multiple scenarios. the data center market is difficult to recover, but intel still has cars, pcs, or robot terminals that may appear in the future to work hard on.

pc is a battlefield that intel cannot lose. in 2023, intel first proposed the concept of aipc and released a processor code-named meteor lake at the end of that year. intel said that this chip integrates something called a neural network processor (npu), which can bring users "energy-efficient ai acceleration and local reasoning experience" on pc.

however, to date, aipc is still just a concept used by major companies to promote their products. according to a more relaxed standard, apple's macbooks using m-series chips are already aipc, because these personal computers are said to be able to run some small large language models, even though apple has not yet released any large language models or made any large-model-based modifications to its computer chips.

similarly, according to relatively loose standards, there are already many chip suppliers in the aipc field, including not only old rivals of intel such as amd, but also new players such as qualcomm and nvidia. at the computex conference held in taipei, china in june this year, qualcomm released the snapdragon x series aipc chip based on the arm architecture, and nvidia released the aipc chip called geforce rtx, and said that asus and msi computers have become its customers.

it is becoming increasingly difficult for intel to hold onto the pc market, and automotive chips are another area where intel has long been planning but may lose opportunities. in 2017, intel spent $15.3 billion to acquire mobileye, the former leader in advanced driving assistance system (adas), focusing on the new computing scenario of autonomous driving. before 2020, smart driving chips were almost monopolized by the company. but soon after the acquisition, mobileye's customers began to churn out, first tesla, then nio, li auto, and then traditional car companies such as bmw and audi.

the reason why these customers left mobileye is simple. like its parent company intel, it adheres to a vertical integration model: since its inception, mobileye has provided car companies with a packaged solution of chips plus autonomous driving systems. to use its autonomous driving solutions, oems can only purchase its chips and algorithms at the same time, and cannot independently iterate the autonomous driving system according to specific needs. on the contrary, nvidia provides car companies with a more flexible solution. its autonomous driving system and chips are sold separately. this solution caters to the new trend of car companies that want to widen the gap with their competitors in autonomous driving systems-they only want nvidia's chips, and as for the autonomous driving systems running on the chips, they want to develop them independently. according to data from gaogong intelligent automobile, in the first half of 2024, horizon robotics has surpassed mobileye and ranked first in the market share ranking of intelligent driving computing solutions for chinese independent brand passenger cars.

it seems that intel's war with amd is coming to an end, and it will have more battles to fight next, but the questions of how to choose opponents and how to allocate resources will never become outdated.

"for the past 30 years, the company (intel) has been focused on suppressing amd, which has caused them to ignore important trends in mobility and low power consumption," said hector ruiz, who served as amd ceo from 2002 to 2008. "they should focus more on their customers and the future instead of trying to surpass amd." ruiz's words are also applicable to the next era of artificial intelligence.

-END-