The Shape of 2025: AI Models Remap Competition, EVs Hit 21 Million Sales, and Gene Editing Goes Mainstream

The New AI Landscape: More Models, Higher Stakes

The last several years in artificial intelligence have been, by virtually any measurable standard, extraordinary. OpenAI released GPT-4 in March 2023, the fourth generation in its Generative Pre-trained Transformer series, built on a multimodal architecture that could process both text and image inputs. Standardized testing performance quickly distinguished it: it passed the U.S. Medical Licensing Exam by more than 20 points above the minimum passing threshold without any specialized prompt construction, solving the kind of nuanced reasoning problems that require synthesis across disconnected domains. By April 2024, OpenAI had shipped GPT-4 Turbo with a 128K token context window, enabling a single conversational thread to retain nearly a novel's worth of context before historical information fell out of scope. By the time the GPT-5 engine took over as the foundation of ChatGPT in early 2026, the platform's weekly active user base had passed 900 million. That figure represents a sustained adoption velocity that no technology platform in recent memory has matched. The practical implication is structural: industries that had previously distanced themselves from AI experimentation—regulatory affairs, senior legal practice, clinical medicine—are now deploying production AI workflows because their practitioners are already using the technologies daily.

The Multimodal Winner-Takes-All Contest

The 2025 field is categorically different from the one that existed in 2023. Competent language-only text models are no longer a competitive product; they are entry-level offerings. The standard bearer has moved through vision, speech, and agentic tool use into a genuinely multimodal plane where a model can receive a camera feed of a document room, a spoken question, and a reference spreadsheet, and synthesize a structured response that integrates and cross-references all three inputs simultaneously. Anthropic's Claude became the developer's choice for extended reasoning chains and constrained outputs, winning admiration among coding teams for its reasoning verbosity in explanation and its willingness to refuse harmful requests before they execute. Google DeepMind, leveraging its parent company's more than a billion end-user accounts, threaded Gemini functionality into search, Gmail workspace features, and the Google Cloud console in a way that required no individual user friction to adopt at scale. This embedding strategy matters enormously: rather than forcing users to learn a new interface, the AI arrived in workflows they were already running.

The Open-Source Wave

Amid proprietary platform growth, open-source AI made its most credible competitive attack yet. Mistral AI's open-licensed models and Meta's Llama ecosystem produced architectures that performed within striking distance of equivalent model sizes under commercial subscription at a zero marginal cost. For organizations in finance, healthcare, defense, and public-sector governance—where data leaving the enterprise perimeter is functionally unacceptable—the evaluation question changed. It is no longer whether open-source AI can match closed-source capability; it is whether the architecture that overfits to the enterprise's operational context offers enough differentiation to justify maintaining a proprietary inference fork.

Inference Economics Turns Competitive

Perhaps the most consequential structural shift of 2024–2025 was the collapse in inference pricing. Model-distillation techniques, where a large teacher model's knowledge is compressed into a much smaller student model sufficient for most production use-cases, brought token-call costs down by a factor of ten in some provider-pair comparisons. Large-model-hosting architecture co-design optimized for dense tensor operations on purpose-built silicon hosting nodes, further reducing the FLOPS-per-token ratio. The practical effect: a startup building a product with AI at the core could in 2025 call the same model family a million times for the cost of a few thousand calls in early 2023.

Agentic Workflows and the Leap to Autonomy

The final frontier of the current AI product cycle is the transition from responsive to proactive. Early 2025 saw the first broadly available agentic workflow products: systems that can be given a high-level objective—'scout acquisition targets in European SaaS', or 'audit Q2 vendor invoices for anomalies'—and that execute the full multi-step chain of tool invocations, conditional branching, and information synthesis before returning a synthesized deliverable. The difference between that capability and the conversational UX we became familiar with in 2023 and 2024 is categorical; one is a cooperative assistant, the other is a junior who can handle an assignment. The revenue impact for knowledge-work enterprises will become visible in 2026 and 2027 as agent deployment matures.

The Electric-Vehicle Revolution Crosses Its Tipping Point

The statistic most often cited in electric-car coverage toward the end of 2024 and early 2025—21 million plug-in electric vehicles sold globally in 2025—merits unpacking to appreciate its full importance. It is not a 'slow adoption curve' statistic. It is a tipping-point marker. At 21 million units, battery-electric vehicle sales constituted more than a quarter of all new vehicle deliveries globally, and the year-over-year growth rate remained at 20 percent, in a segment where the historical growth curve had already been steep. Together these signals confirm that conventional internal-combustion vehicle sales have already peaked and are now in a structural decline.

For the next decade, the electric vehicle conversation will not be about whether the technology dominates; it will be about which geography, manufacturer, and standardized architecture captures the biggest share of the downslope. China's position as the world's largest EV fleet holder is not yet contested: cumulative electric vehicle sales, including heavy commercial vehicles, buses, and sanitation machines manufactured within China, exceeded 22 million units through late 2024. The United States and the European Union have converged on a similar moment of cost-structure parity: as of 2020, the total cost of ownership of new electric vehicles has been lower than equivalent internal-combustion vehicles across both geographies when fuel savings and reduced maintenance are properly capitalized across ownership horizons.

Mass Market Acceptance and the Tesla Baseline

The 2023 milestone where the Tesla Model Y became the best-selling vehicle of any powertrain type globally is the single clearest signal that electric vehicles have crossed from a niche luxury adoption into a mainstream product category. A crossover SUV built on a lithium-ion battery pack did not displace gasoline-engined tractors with volume pricing or a century of combustion-engine infrastructure by accident; it won because the product category it replaced now has category-defining disadvantages that consumers registered and consistently cited as purchase objections.

The Tesla Model 3 established the all-time sales record for electric vehicles and the first EV to cross the million-unit threshold in global cumulative deliveries in June 2021. Together the Model 3 and the Model Y have stretched the distribution curve of EV purchasers and pulled forward the market entry dates for vehicle architectures previously considered mid-market. The knock-on competitive effect of those volumes spread across every global OEM: the consistent-performance and cost floor lifted by Tesla's scale raised the bar that every competing internal-combustion platform and EV platform had to clear.

Tesla's 2025 production year saw the company produce approximately 1.66 million vehicles, sustaining its billion-dollar operating-income base and generating revenue in the US$94.83 billion range for the fiscal year. In late 2024 and through mid-2025, Tesla re-entered trillion-dollar market-capitalization status on multiple occasions—a useful proxy for the scale of investor conviction in EV's long-run earnings. What is important to track going forward is whether Tesla's advance accelerates further into high-volume vehicle delivery, reasserts technological differentiation through Full Self-Driving claims, or faces competition from Chinese manufacturers who have demonstrated that at the right cost-structure point they can sell at price points European and American OEMs have been reluctant to attempt.

Charging Infrastructure: Cole's Third Law

Recharging accessibility is the single most underappreciated variable in EV adoption numbers outside of vehicle-level cost. The charging station ecosystem in most Western geographies did not grow in proportion to vehicle adoption, creating localized supply constraints. Tesla's decision to open and standardize its Supercharger architecture in 2023 following regulatory pressure and to integrate the North American Charging Standard into other manufacturers' products significantly improved the probability of long-distance EV travel on continental-scale routes. The Supercharger network had grown to over 7,700 stations globally, providing a foundation that competitive private networks are working to match. Battery-swapping technology—already implemented at scale in China—merits particular attention in markets where the supply of apartment-dwelling drivers for whom home-charging infrastructure is structurally unavailable is large.

Autonomy: An Edge Case That Remains

The Full Self-Driving feature announced in Tesla's investor event in 2024 and continuing toward year-end 2025 sits in a peculiar conceptual space. On highway segments and structured urban corridors with good visual map data, Tesla vehicles using the FSD beta demonstrated a level of navigational competence that represents a meaningful departure from the adaptive cruise-control features of earlier-generation cars. Mixed pedestrian situations—school zones, markets where two-wheeled traffic shares lanes with four-wheeled, large-vehicle construction redirects—remain the reliable edge conditions that degrade system reliability in ways that human drivers are trained to handle. The autonomous-driving effort of 2025 cannot be judged on whether Level 5 full-vehicle autonomy, under every possible driving context, was reached. By any fair standard, it was not. The more relevant standard for commercial deployment is whether the system can, in the geographic markets where it operates, gain airworthiness certification equivalent to that demanded of a commercial aircraft's automated flight systems.

Hardware reductionism is one area where significant progress was made. Lidar sensors, which dominated advanced sensor budgets in 2019–2022, fell to commodity pricing. High-resolution camera arrays with embedded local processing are now shipping in sufficient density to make radar and lidar elegant rather than essential to a deployment stack. Parameter-efficient fine-tuning and synthetic data cycles that allow the perception and planning models to learn from the accumulated fleet miles are extending the functional competence of existing hardware stacks far more quickly than the hardware upgrade cycle alone would have achieved.

Biotech at Definite Inflection: Gene Editing Goes From Lab Curiosity to Launched Medicine

The biopharmaceutical industry was for decades the slowest of all big industries at adopting advanced computation. Drug development cycles expanded to average twelve to fifteen years with costs in the billions per approved molecule. Then mRNA vaccine platform-readiness produced a Gravity Well Effect that collapsed the entire debate: there was now a proof of concept for a platform-delivery approach to therapeutics that could be initialized and scaled in months rather than years, and that could be redirected from one molecular target to another on timelines that were previously unthinkable. The COVID-19 pandemic accelerated a shift that was already structurally positioned; the practice of medicine-paying rationalized its acceptance of the conceptual model and turned the platform-capable future from speculation into a funded operational category.

CRISPR-Cas9: The Technology That Won a Nobel Prize and a Drug Pipeline

CRISPR-Cas9—the short form of Clustered Regularly Interspaced Short Palindromic Repeats, a bacterial antiviral defense system turned molecular scalpel—represents the most consequential breakthrough in molecular biology since the structure of DNA was characterized by Watson and Crick in 1953. Jennifer Doudna of the University of California, Berkeley, and Emmanuelle Charpentier of the Max Planck Unit for the Science of Pathogens jointly received the 2020 Nobel Prize in Chemistry for developing CRISPR gene editing. Their work established that the Cas9 nuclease enzyme, when programmed with a synthetic guide RNA sequence, could be directed to make a double-stranded break in any targeted stretch of DNA—opening a precise, controllable, and generalizable pathway for genome modification. The mechanism works as follows: a guide RNA, programmed to be complementary to a specific DNA sequence, is introduced into a cell alongside a Cas9 enzyme complex. The Cas9-GRNA complex locates the target sequence and makes a clean double-strand cut at that precise location. The cell's natural DNA-repair machinery then goes to work. In one repair pathway—non-homologous end-joining—the broken ends are joined back together, often with random insertions or deletions that can effectively silence a disabled gene. In the homology-directed repair pathway, the cell can be provided with a template DNA sequence that replaces the original, allowing for a genuinely precise correction rather than a disruptive intervention.

The practical scope of this capability is a function of how many different cell types in which the editing tool operates efficiently. For somatic cells—the non-reproductive, tissue-forming cells that constitute almost all of the body's functional volume—the modifications introduced affect only the cell in which they occur, making the modification a direct, local therapeutic act that does not propagate to future generations. This distinction, legally and ethically manageable in most jurisdictions, is what makes therapeutic somatic CRISPR use clinically and regulatorily tractable, distinguishing it sharply from germline editing applications that persist in the human reproductive line and which remain subject to a near-universal global prohibition.

Casgevy: The First CRISPR Drug and a Global Milestone

The most vivid confirmation of CRISPR's transition from a laboratory technique to a validated therapeutic modality arrived in late 2023, when the UK Medicines and Healthcare products Regulatory Agency approved exagamglogene autotemcel—marketed as Casgevy—for the treatment of sickle cell disease and beta thalassemia. Bahrain gave regulatory approval in December 2023. The US Food and Drug Administration followed on 8 December 2023, conditioning the approval on clinical evidence of durable correction in treated patients. For the estimated 400,000 individuals born with sickle cell disease worldwide and the disproportionate health burden it imposes on sub-Saharan African populations, a single-course curative treatment with measurable durable effect changes a lifetime of clinical crisis management into a corrected baseline. Beta thalassemia, which impairs the production of functional hemoglobin in the bone marrow and requires regular transfusions for survival across its most severe variants, addresses a different—but also geographically stratified—unmet need. The three-country cross-national regulatory approval happened at essentially the same time as one of the strongest conceivable single signals that biotechnology evaluation frameworks had absorbed the new platform-speed paradigm.

The same concern that surrounds cutting-edge technology—off-target effects, unintended mutations, precise adverse-event profiles in long-term follow-up cohorts—remain in active-supervision mode on these therapies. Regulatory frameworks in several geographies explicitly oblige long-term post-approval monitoring registries specifically designed to persist for decades for cohort members born with the treated condition. The ethical structure that governments erected while awaiting the science keeps pace.

The mRNA Platform: From COVID-19 to Cancer and Beyond

mRNA vaccine technology demonstrated at nearly 2.7 billion doses administered globally during the COVID-19 vaccination campaign, proved more than just a rapid-response mechanism. It proved that a platform-delivery approach could be deployed at scale, directed to new molecular targets, and reshaped to address new diseases with a timelines contraction that was not available to conventional small-molecule and therapeutic antibody platforms at any point in their history.

The post-pandemic mRNA pipeline is now forward-loaded with oncology trials. Customizable mRNA cancer vaccines—tailored to an individual patient's tumor mutation profile—are progressing through Phase 2 and Phase 3 clinical trials against melanoma, non-small-cell lung cancer, and colorectal cancer with statistically significant early outcomes. Also under active investigation is the lipid-nanoparticle delivery architecture that is the mRNA platform's enabling technology, being reformulated for active delivery systems for hepatocytes, neurological tissue, and other targets that have historically been inaccessible or poorly accessible using conventional small-molecule pharmacokinetics.

CAR-T Cell Therapy Rewriting the Cancer-Treatment Paradigm

CAR-T cell therapy — an adoptive cell transfer approach that equips a patient's own T cells with synthetic targeting molecules for specific cancer antigens — has, in its best-performing specifications, achieved complete remission rates that conventional cytoreductive chemotherapy regimen have rarely approached. Amyloid-lowering antibodies approved in late 2024 and early 2025 for Alzheimer's disease represent the antibody-as-platform trend reaching neurodegenerative diseases: the same bioengineering principles that produced effective antibody treatments for inflammatory conditions, infectious diseases, and cancer are now delivering pharmacological options for diseases that had, until very recently, nearly no effective treatment.

The Convergence Thread

No domain listed here is truly independent of the others. AI machine learning systems have already reduced early-stage drug discovery cycle times by accelerating the simulation of molecular interaction and hit identification stages that previously required physical library-screening through labor-intensive wet lab experiments. Autonomous-driving capacity ultimately depends on the compute architecture and model training pipeline capabilities that general AI platform providers commercialize. Sustainable-electric-infrastructure fast-charging roll-out rates depend on the same semiconductor and battery density improvements that are simultaneously feeding consumer electronics and renewable energy storage.

The institutional question facing societies over the next half-decade is not whether these compounding trajectories are desirable on any absolute quality-of-life metric; they broadly are, across employment, health access, connectability, and environmental externalities. The harder question is how societies accelerate the equitable distribution of technology access as quickly as the technology accelerates. The AI platforms currently replacing knowledge workers by function are doing so faster than social-protection institutions are retooling for the structural adjustment. The electrified-transportation supply chain is shifting the geography of automotive employment and manufacturing in ways that national industrial strategies must manage. The gene-editing and personalized-medicine pipelines have raised the baseline cost of curative medicine in ways that threaten to deepen healthcare access inequality unless deliberate payment-system design accompanies approval.

None of those dilemmas are reasons to slow the technology. They are reasons to engage the policy and institutional dimension as aggressively as the innovation dimension. The compounding cascade is real and accelerating; what happens to the distribution of its benefits is not yet written.