As the curtain falls on 2025, a series of landmark business events have left a clear and profound mark on the year. The AI race has heated up comprehensively. The White House ignited "Stargate," a $500 billion gamble on America's AI infrastructure; CoreWeave placed computing power leasing on the capital market's scale; OpenAI, though not publicly listed, leveraged orders and narratives to sway trillion-dollar market cap fluctuations, becoming a "shadow giant" of the market. NVIDIA was crowned the world's first $5 trillion company, invested in Intel, and secured Groq's inference capabilities, consolidating its computing throne through alliances and defenses. Google, driven by both TPUs and Gemini, launched a direct assault on AI pricing power. On the capital stage, the 95-year-old Warren Buffett penned his "final letter," signaling the end of an era. Looking further, Elon Musk used 155 launches to push spaceflight into an industrial rhythm. Meanwhile, Europe slammed on the brakes as Germany rewrote its combustion engine ban, buying time for traditional industries. In Europe, Germany's sudden reversal of the 2035 internal combustion engine ban, allowing a reprieve by permitting e-fuels, exposed the tension between aggressive transitions and market realities. The old order loosened, and new forces rose. Alliances, games of strategy, high-stakes bets, and reversals formed the most authentic business backdrop of 2025. As the year ends, we review the ten most representative global business events of 2025, looking back at how they reshaped the tech landscape, capital logic, and the direction of our times.
The White House ignited "Stargate": a $500 billion bet on America's AI foundation. In January 2025, "Stargate" was unveiled with great fanfare at the White House. OpenAI, SoftBank, and Oracle announced a rolling four-year investment of $500 billion, with an immediate $100 billion injection, to build 20 super-large-scale AI data centers across the US, creating the "largest AI infrastructure project in history." Masayoshi Son assumed the role of chairman, exuding immense momentum. However, reality began to cool just two months later. In March, construction started on the Abilene, Texas base, planned for 64,000 NVIDIA GB200 chips, with an initial phase of 16,000. But public disagreements emerged between SoftBank and OpenAI over equity ratios, data sovereignty, and funding schedules, revealing cracks in the partnership. By May, multiple media outlets reported the project was "stalled." The full $500 billion was not secured, and the planned summer installation milestone was delayed. To secure computing power for ChatGPT, OpenAI bypassed SoftBank, signing contracts directly with third-party data centers, forcing the "alliance project" to yield to practical needs. In July, the goals were further scaled back. The company revised its 2025 target to "completing a small pilot center in Ohio by year-end," a stark contrast to the "immediate $100 billion" promise of January, with "shell-preservation progress" becoming the new tone. On October 30, the project announced a new 1GW campus in Saline, Michigan, as a symbolic fulfillment of the "total 4.5GW plan," expected to break ground in early 2026. The scale remained but was significantly below the initial blueprint. In November, financing and supply chain supplementation proceeded simultaneously: Blue Owl provided $3 billion in equity and an $18 billion syndicated loan; in the same month, SoftBank acquired server chip company Ampere for $6.5 billion in cash, effectively "buying a ticket" for the subsequent supply chain. Entering December, Abilene remained the only site with substantive progress. Four server buildings were topped out, with mechanical, electrical, and cooling systems being installed, on track to deploy 400,000 GPUs by mid-2026, potentially becoming the "largest known single AI cluster." However, sites in Arizona, California, Florida, and others still lacked formal agreements. Concurrently, risk hedging accelerated. Oracle delayed construction of at least one data center and sold its Ampere shares for $2.7 billion, reducing its unilateral exposure. SoftBank, facing obstacles in external financing, sold some NVIDIA shares to raise cash, indicating the $500 billion was far from fully assembled. From an "epic announcement" to "point-specific landing," Stargate did not fail but was significantly downsized. The conclusion became clearer: Abilene is poised to become a US AI computing landmark by mid-2026, but the original vision of "$500 billion, 20 campuses" has been reshaped by博弈, financing, and uncertainty. The true test will come after Abilene goes live, with operational results and a new round of capital providing the answers.
From a "GPU hoarding business" to public pricing: CoreWeave (CRWV.US) puts AI computing power on the capital market's scale. On March 28, CoreWeave listed on the Nasdaq at $40 per share, with a valuation of approximately $23 billion, raising $1.5 billion. It opened at $39, and over the next four months, its stock price surged to a high of $187, briefly pushing its market cap close to $90 billion. This was not just the IPO of one company; it was the first time "AI computing power leasing" was priced by the public market. CoreWeave's origins were humble. In 2017, it was an Ethereum mining operation. After the crypto winter, the company repurposed its 250,000 NVIDIA GPUs into cloud assets, transforming into a "pure GPU cloud," eschewing CPUs and general IaaS to focus on one thing: hoarding GPUs and renting them out. Its revenue formula was highly simple: priced per GPU, locked in long-term. Major clients signed multi-billion dollar contracts for 3 to 6 years upfront, while smaller clients paid by the hour, creating highly visible cash flows. Eleven months before its IPO, CoreWeave signed an $11.9 billion, 5-year lease with OpenAI; in September post-IPO, it added another $6.5 billion, bringing the total to $22.4 billion. Around the same time, Meta secured orders worth up to $14.2 billion, and NVIDIA committed to $6.3 billion in "take-or-pay" agreements for unsold capacity, backing the demand side. These three major orders pushed the backlog to $55.6 billion, 29 times its 2024 revenue. The valuation was rapidly inflated: at the IPO price, CoreWeave traded at 12 times its 2024 revenue P/S; at its peak market cap, the P/S ratio approached 47x, higher than NVIDIA's at the time, earning it the short-seller label of "core of the AI bubble." Behind the controversy, the industry's coordinates were rewritten. CoreWeave proved that "compute middlemen" could list independently and gain institutional recognition, prompting peers like Lambda and Crusoe to initiate their own IPO preparations. Its structure of "long-term leases + GPU collateral financing" also provided banks with a quantifiable model, promoting pilot GPU leasing ABS and lowering the financing threshold for the AI industry chain. More importantly, it changed perceptions. Public market volatility inversely educated clients: computing power was no longer just a hard-to-get "hard currency" but an asset that could be compared for lease-to-buy ratios and flexibly transferred. The industry was moving from "GPU hoarding panic" to "on-demand configuration." From a mining farm to the main board, CoreWeave wrote "GPU as asset, compute as service" into the valuation model. $23 billion was just the starting point; what was truly being priced was a new category of capital expenditure.
NVIDIA (NVDA.US) invested $5 billion in its old rival Intel (INTC.US). NVIDIA and Intel, long-time adversaries, suddenly chose to join hands. On September 18, NVIDIA announced it would invest $5 billion in Intel and jointly develop chips for PCs and data centers. According to the agreement, NVIDIA purchased Intel common stock at $23.28 per share. The market reacted swiftly after the announcement. Intel's US pre-market stock surged up to 30%, while rival AMD plunged over 4%. For Intel, this was a timely lifeline. In recent years, it had been losing ground in the high-performance chip market, struggling to fund hefty investments in advanced processes with its own cash flow. The company had already secured about 10% ownership support from the US government, a $2 billion strategic investment from SoftBank, and accelerated financing through asset sales. NVIDIA's involvement further stabilized its funding chain. For NVIDIA, this was a high-impact, low-cost maneuver. $5 billion was not a heavy burden relative to its size, but it opened doors to the x86 ecosystem and broader CPU synergy. The core of their collaboration targeted two key battlegrounds: PCs and data centers. Intel will integrate NVIDIA's graphics processing technology into its next-generation PC chips to enhance competitiveness against AMD; in data centers, Intel will provide general-purpose processors for AI clusters built on NVIDIA hardware, complementing the shortcomings of accelerator chips in general computing. NVIDIA CEO Jensen Huang stated this partnership would combine NVIDIA's AI and accelerated computing stack with Intel's CPUs and vast x86 ecosystem, laying the foundation for the next computing era. Intel CEO Pat Gelsinger emphasized that x86 remains the cornerstone of modern computing and that the company would continue to drive innovation. Both sides also indicated the collaboration would not alter their independent strategies, with NVIDIA continuing to use Arm technology to design its own processors. This turn of events seemed almost inevitable after a period of shifting fortunes. From rivals to partners, this $5 billion bet was not just a capital transaction but a significant signal in the reshaping of the industrial landscape: in the new AI-dominated computing era, closed defenses are unsustainable, and alliances between the strong are becoming the new rule of survival.
The shadow giant's roller coaster: OpenAI stirs the 2025 AI capital narrative. In 2025, OpenAI did not go public, yet it became the strongest "sentiment engine" in the capital markets. Financing, orders, and executive comments were continuously magnified across the entire AI industry chain,演绎出一轮轮过山车行情. From January to April, the story was ignited by infrastructure. OpenAI, partnering with SoftBank and Oracle, launched the $500 billion "Stargate" compute alliance, with Trump endorsing it at the White House. Oracle's stock soared in a single day, SoftBank surged in tandem, and the market quickly cast OpenAI as the "standard-bearer of AI infrastructure," leading to a collective re-rating of compute, server, and chip stocks. From May to August, cracks appeared in the halo. The August release of GPT-5 delivered mediocre performance, being outperformed by Google's Gemini 3.0 and xAI's Grok on multiple metrics. Simultaneously, the mismatch between an "approximate $20 billion annualized revenue" and a "valuation exceeding $500 billion" was frequently mentioned, with the "AI revolution" narrative beginning to be replaced by "AI bubble," noticeably cooling the rally in concept stocks. In September, OpenAI announced its first consumer-grade AI hardware would be manufactured by Luxshare Precision, sending Luxshare's stock limit-up at the open and lifting the "Apple supply chain" collectively; in the same month, Oracle's remaining performance obligations rose to $455 billion, pushing its stock to new highs and pulling sentiment back up. In October, intensified financial concerns roiled market sentiment. SoftBank recorded $14.6 billion in unrealized gains from its OpenAI stake, reporting record quarterly net profit and sending its stock to a peak; but OpenAI's acquisition of financial AI app Roi and deep ties with AMD also led outsiders to calculate its cash burn and financing intensity. For the first time, CDS markets saw default swap trades for several tech giants, bringing debt risks to the forefront. On November 7, a sentiment "meltdown" occurred. The CFO's comment that "federal government backing for chip financing might be needed" was interpreted as "even the strongest AI company is short on cash." That evening, the six major US tech giants and compute suppliers collectively lost about $500 billion in market cap. CEO Sam Altman hastily posted a clarification that "no government guarantee is needed," but it only prompted a limited rebound, as the market entered a "trust deficit." In December, the competitive landscape reversed. Google's Gemini gained sustained positive口碑, with the Alphabet ecosystem chain doubling within the year; in contrast, the "OpenAI camp," including Oracle and SoftBank, fell about 40% from their highs, shedding over $100 billion in market cap, turning the "OpenAI concept" from a premium label into a risk exposure. Looking back on the year, OpenAI's "stock price effect" was first pushed to a boiling point by grand infrastructure and hardware narratives, then continuously stirred by model competitiveness, debt concerns, and verbal missteps. The capital logic had shifted from "changing the world" to "delivering cash flow." Whether it can transition its roughly $20 billion revenue towards more sustainable growth will determine whether the AI sector moves towards maturity in 2026 or continues its cycle of high valuation and high volatility.
NVIDIA becomes the world's first "$5 Trillion Company." On October 30, NVIDIA's stock rose about 3% to $207.16, lifting its market capitalization to $5.03 trillion, making it the first company to breach this threshold. This figure exceeded the combined market cap of AMD, Arm, ASML, Broadcom, Intel, Lam Research, Qualcomm, and TSMC, and was larger than the entire sectors of Utilities, Industrials, and Consumer Staples within the S&P 500. Over the past six months, NVIDIA's stock had climbed approximately 90%. Its current size even surpassed the combined market capitalization of the major stock indices of Germany, France, and Italy. Three years prior, before ChatGPT's debut, NVIDIA's market cap was only about $400 billion. Subsequently, demand for GPUs needed to train and run large models surged, pushing its market cap past $1 trillion within months, and accelerating from there: reaching $2 trillion in February 2024, $3 trillion in June, $4 trillion in July 2025, and breaking $5 trillion in October. Demand was the core driver of this leap. CFRA Research Senior Vice President Angelo Zino stated bluntly: "They are the bedrock of the entire AI trade." NVIDIA disclosed that it had shipped 6 million units of its Blackwell chip, released last year, with another 14 million units on order. CEO Jensen Huang, at the GTC conference, projected total sales of $500 billion over the next five quarters. But amid the frenzy, skepticism also grew. Some investors drew parallels between the current AI stock trajectory and the dot-com bubble: companies investing hundreds of billions, taking on more debt, while real revenue remained limited. Valuation also faced pressure, with NVIDIA's stock trading at about 33 times next year's expected earnings, higher than the S&P 500's average of about 24 times. From $400 billion to $5 trillion, NVIDIA completed a "10-bagger" leap in three years. It not only reshaped its own destiny but also defined the computing coordinates of the AI era. Yet, with its market cap soaring into the trillions, the growth story must continually deliver to justify this most dazzling crown.
Buffett's "final letter": I was "just plain lucky," but "Father Time" has caught up, I will "keep quiet." The 95-year-old Warren Buffett wrote his final annual shareholder letter on November 10, 2025. The phrase "I'm going quiet" declared his intention to step down as CEO at year-end and withdraw from daily management. The investor who had led Berkshire Hathaway for nearly six decades formally entered his final act. The baton was passed to new CEO Greg Abel. Buffett praised him highly for his excellent management, diligence, pragmatism, and sincere communication, even joking that he hoped Abel would "keep going forever." Henceforth, the annual letter would be penned by others, but Buffett promised to still write to shareholders every Thanksgiving, maintaining the emotional bond with Berkshire. The most moving part of this "farewell letter" was his reflection on life. At 95, he thanked his "incredible luck": being born in the US in 1930, blessed with health and the红利 of the era. A near-fatal appendicitis in 1938, where he reportedly asked a nun for fingerprints in the hospital, fantasizing about solving cases for the FBI. Mentioning Charlie Munger and other old friends, he quipped that "perhaps there's something in the water in Omaha." Now slowing down and finding reading difficult, he still insisted on going to the office weekly, occasionally still sparking good ideas. At this moment of departure, he again offered business maxims, directly targeting corporate greed. Buffett criticized executive pay disclosures for triggering a "richer-than-thou" contest: "What troubles very wealthy CEOs is often that others are wealthier." He cautioned Berkshire to avoid leaders eager to retire at 65, pursue "conspicuous wealth," or attempt to build a "dynasty." His long-termism also stood in contrast to the current speculative fervor. Starting with investing in a struggling Berkshire in 1962, he built it into a business empire spanning insurance, manufacturing, utilities, railroads, and brands like Dairy Queen and Fruit of the Loom, emphasizing the need to avoid paths that could reduce the company to a "beggar." One letter, concluding a lifetime. Stepping away from power, adhering to principles, giving back to society. For the market, this was a historic turning point for Berkshire; for investors, it was a final demonstration built on the bones of data and the soul of reason.
After 155 launches, Musk pushes spaceflight into the "Industrial Age." In 2025, SpaceX redefined the meaning of "spaceflight scale" with a launch tempo so dense it bordered on疯狂. By November 22, Falcon 9 had completed 150 orbital-class missions, plus 5 Starship test flights, totaling 155 launches for the year—an average of one rocket launch every 2 to 3 days. Cape Canaveral alone accounted for 94 launches, about 40% of the global total. For the first time, spaceflight exhibited the rhythm of an industrial assembly line. Musk's core advantage was the "re-evolution" of reusability. Booster B1067 achieved 31 flights with a single vehicle, setting a world record; over 20 active boosters in the fleet had surpassed 10 flights, with certification targets being pushed to 40 flights. For the year, there were 532 first-stage recoveries and 454 re-flights, with a success rate exceeding 97%. The concurrent reuse of fairings and grid fins pushed the marginal cost per launch down to the $150,000-$200,000 range, transforming rockets from "expensive equipment" into "recyclable assets." Starship represented the longer-term bet. In five comprehensive test flights during the year, the latest achieved a complete recovery of the second stage upon re-entry; the Block 2 hull and heat shield tiles underwent over 300 ground thermal cycle tests. Launch Complex 39A at Kennedy Space Center was being converted into a "Starship-dedicated" pad, paving the way for the first East Coast launch in 2026 and subsequent HLS and GTO missions. An industry深度研究报告 released by Sinolink Securities on December 10 indicated that SpaceX's essence is not that of a traditional aerospace manufacturer, but rather a "monopoly provider of space logistics and infrastructure that applies first principles to the extreme." The report noted that its seemingly insurmountable industry barriers stem not from a single technological breakthrough but from the deep integration of moats across three dimensions: cost, manufacturing, and customers. Currently, SpaceX has formally initiated the process of selecting investment banks, marking the most substantial step yet towards an IPO for this commercial space giant. If progress is smooth, this could become one of the largest IPOs in the capital markets in recent years. This is not just Musk's victory; it signifies that commercial spaceflight is transitioning from "engineering marvel" to "scale industry." The foundation of the great space age is being solidified by rocket after rocket flying again.
Banning the ban: Germany buys time for the combustion engine. After being in effect for over two years, the EU's climate agenda "crown jewel," the 2035 ban on new internal combustion engine cars, was shattered by forceful intervention from Berlin. The once unshakeable "100% zero-emission" target was urgently amended to "90% reduction." According to the latest方案 disclosed on the 23rd, the remaining 10% gap could be filled using e-fuels or biofuels. This meant that, as long as filled with "green fuel," internal combustion engines and plug-in hybrids would retain legal road rights post-2035. Behind this reversal lay a disconnect between冷酷的市场数据 and aggressive emission reduction targets: Market Gap: The 2025 compliance requirement for automakers' carbon reduction demanded a 25% share for electric vehicles, but the actual share was only 16.4%. Adoption Divide: EV sales accounted for 35% in the Netherlands but only 8% in Spain, making a "full electric transition" unrealistic due to uneven infrastructure. Fine Threshold: Maintaining the original plan would expose European automakers to fines amounting to hundreds of millions of euros. German Chancellor Merz and European People's Party (EPP) President Weber joined forces to end the "pure electric ideology." Weber hailed the move as a "major achievement in reconciling climate goals with market realities." Through intensive lobbying, giants like Mercedes-Benz and Volkswagen ultimately pressured the von der Leyen government to "relent" under political pressure. In the short term, Germany leveraged its strong trade leverage to secure a survival window for its domestic ICE supply chain; but in the long term, it remains doubtful whether this "reprieve" can withstand the迭代速度 of Chinese automakers. As German Finance Minister Christian Lindner warned: "If we cling to old dreams, the future situation will only become more difficult."
Google's rise: Dual-driven by chips and models, charging towards the global number one spot. Alphabet is turning a technological evolution into a market cap race. In 2025, propelled by the AI wave, its stock rebounded over 60%, lifting its market capitalization to $3.8 trillion and elevating it to the world's third most valuable company. This bet's core lies not in advertising, but in the dual-drive of "chip + model." On one end is hardware. Google is betting on its self-developed TPU, partnering with Meta to advance a plan codenamed "TorchTPU," targeting NVIDIA's most坚固的护城河—CUDA. By enhancing TPU compatibility with PyTorch, Google aims to "unlock" developers from the GPU ecosystem. More importantly, TPUs are no longer just for its own cloud but are beginning to be sold directly to external customers. Under the coordination of new AI infrastructure head Amin Vahdat, Google is pushing its internal capabilities to the market, building a compute option independent of the NVIDIA system. On the other end is the model. Gemini 3 Flash is positioned as the efficient "workhorse model": three times faster than the previous generation Gemini 2.5 Pro, at just a quarter of the cost of Gemini 3 Pro; it even outperformed the more expensive version in the SWE-bench Verified programming test. Google quickly set it as the default engine for the Gemini App and search AI mode, leveraging its distribution advantage for scale, and using scale to gain data and stickiness. Behind this lies a strategic重构: using TPUs to reduce compute costs, using Gemini to compress model prices, and then using search and app entry points to complete the distribution loop. Google's aim is not just to catch up with OpenAI, but to fundamentally shake NVIDIA's pricing power in AI infrastructure. The trend is clear. AI competition is shifting from "who is stronger" to "who is cheaper, who is easier to use, who is easier to migrate." If the hardware-software synergy continues to gain traction, Alphabet could become not just a cloud provider and model maker, but potentially the rule-setter for the next-generation AI platform. The battle for market cap has evolved from capital博弈 into a head-on collision of technological routes. Google is charging towards the throne.
Buying brains, not the company: NVIDIA's "quasi-acquisition" of Groq in the inference war. The long-rumored并购靴子 finally landed. On December 25, NVIDIA spent approximately $20 billion to lock in the core capabilities of AI chip unicorn Groq through a "technology licensing + talent acquisition" deal. It did not acquire the company's equity but brought the most critical technology and people into its fold, formally launching a defensive battle centered on "inference." NVIDIA will obtain a non-exclusive license to Groq's low-latency inference technology and bring founder and CEO Jonathan Ross, President Sunny Madra, and the core engineering team onboard. Groq will continue to operate independently, with former CFO Simon Edwards becoming CEO, and the GroqCloud business continuing as usual. This was a "quasi-acquisition" transaction: circumventing antitrust scrutiny while substantially weakening a potential rival. The $20 billion price tag was about three times Groq's $6.9 billion valuation from September last year, reflecting NVIDIA's urgency. The focus of AI computing is shifting from training to inference. Whoever can "run models more cheaply" will determine commercialization costs and profit margins. While GPUs are unmatched in general computing, they face正面挑战 from custom architectures in low latency and energy efficiency. Groq's visionary, Ross, was an early core member of Google's TPU team. The LPU architecture he championed is designed specifically for inference, trading embedded memory for deterministic execution and lower power consumption. For NVIDIA, it's acquiring not just a solution, but a mature technology roadmap. For Groq, this is both a highlight and a reality. Its valuation jumped from $2.8 billion to $6.9 billion within a year, but breaking through independently became increasingly difficult under the pressure of NVIDIA's CUDA ecosystem and in-house development by giants. Being "co-opted" offered a payout and survival. The trend is clear: AI is entering the inference era, and compute competition is moving from scale to efficiency. By buying key capabilities instead of the company, NVIDIA's goal is singular—to ensure the pricing power for the next phase remains firmly in its own hands.
